跳到正文
原文
Nature 健康护理主题·· 16 天前AI 评分49

ResNet-GA:遗传算法驱动的ResNet优化框架

ResNet-GA: genetic algorithm-driven optimization of ResNet for automated dental caries detection in medical imaging

AI 导读

该研究提出了一种结合遗传算法优化ResNet-50的框架,用于牙科龋齿检测。通过自动搜索最佳卷积配置,该框架提高了模型的准确性和可靠性,同时减少了对手动调参的依赖。

正文

Introduction

The rapid evolution of computer systems and software has paved the way for artificial intelligence (AI) to become deeply integrated in various aspects of everyday life. AI has significantly influenced industries such as construction, transportation, and manufacturing, as well as sectors such as business intelligence, education, and health care. Machine learning, a core component of AI, allows systems to learn and adapt based on experience without requiring explicit programming. Specifically, it focuses on creating algorithms that can access and analyze data to allow them to learn independently. The main objectives of machine learning include improving the accuracy of the solution, expanding the capabilities of problem solving, improving the cost efficiency of generating solutions, and optimizing resources through a streamlined knowledge representation1.

Deep learning, a subfield of machine learning, has emerged as a powerful tool for data analysis and predictions2. Inspired by the structure of the human brain, deep learning algorithms utilize artificial neural networks to process large data sets. These networks consist of interconnected layers, each performing increasingly complex computations on input data. This layered architecture allows deep learning models to learn intricate patterns and relationships within the data, leading to superior performance in various tasks such as image recognition, natural language processing, and speech recognition. Deep learning has achieved notable success in a wide range of fields, including healthcare, autonomous driving, and finance3.

  • Automatic translation: Deep learning models can translate text between languages with remarkable accuracy, even capturing nuances and context4.

  • Image recognition: Deep learning algorithms, particularly Convolutional Neural Networks, excel at identifying and classifying objects within images. This capability forms the basis for numerous advanced applications, including facial recognition, medical image analysis, and self-driving cars5.

  • Bioinformatics: Deep learning plays a crucial role in analyzing biological data, such as DNA sequences and genomic information. It contributes to drug discovery, personalized medicine, preventive healthcare, and gene therapy6.

Bioinformatics, an interdisciplinary field that integrates biology and information technology, plays a crucial role in analyzing molecular sequences and genomic data. Using modern computational tools such as cloud computing, statistics, and mathematics, bioinformatics enables the application of sophisticated techniques, such as pattern recognition, machine learning, and molecular modeling. The significance of bioinformatics has been particularly evident in medicine, where sequencing the human genome has provided profound insights into the genetic basis of various diseases (78;).

A prime example of the application of bioinformatics in medical practice is the detection and treatment of dental caries, a prevalent oral disease that involves progressive decay of tooth structures. This condition, often caused by oral bacteria, can cause significant dental problems, including pain, infection, and tooth loss, if left untreated. Preventive measures, such as proper hygiene and a balanced diet, are critical to mitigate the risk of caries. However, diagnosing and classifying dental caries remain challenging and time-consuming processes, highlighting the need for more efficient diagnostic methods9.

This challenge mirrors other medical diagnostic tasks, such as brain tumor detection, where identifying and classifying tumors from medical images is equally complex. Fortunately, the advent of deep learning has provided promising solutions for automating medical image analysis. Using deep learning models, healthcare professionals can streamline the process of diagnosis and treatment planning, improve the accuracy of risk assessments, and enable earlier interventions for better patient outcomes.

The success of deep learning in general image analysis tasks has paved the way for its application in the detection of dental caries. Techniques such as convolutional neural networks (CNN) have been used to detect and classify dental caries, showing substantial promise in automating the diagnostic process (1011;). These models, particularly the ResNet architecture, have proven effective in achieving high accuracy in medical-image analysis.

However, the performance of deep learning models, especially complex architectures such as ResNet, is heavily based on the proper optimization of hyperparameters. Parameters such as learning rate, number of layers, and number of neurons per layer significantly impact the performance of the model. Incorrect hyperparameter configurations can lead to suboptimal results, making their optimization a critical step in model development. Studies have shown that optimizing hyperparameters using techniques such as random search can improve model accuracy, particularly in medical applications12.

Among the various optimization techniques, genetic algorithms (GAs) stand out owing to their ability to efficiently explore high-dimensional search spaces to identify optimal hyperparameter configurations. Inspired by natural selection, GAs evolve populations of candidate solutions and have been shown in several studies to outperform traditional methods, such as grid search and random search7. In the context of dental caries detection, applying GAs to optimize the hyperparameters of deep neural networks has the potential to improve model accuracy, reduce false positives and negatives, and ultimately enhance diagnostic reliability (813;).

Despite recent progress in deep-learning-based dental caries detection, several important gaps remain. First, many existing models rely on manually selected hyperparameters, which makes model development time-consuming, experience-dependent, and difficult to reproduce. Second, conventional hyperparameter search strategies, such as grid search and random search, can be computationally expensive and inefficient for deep architectures. Third, dental images often present practical challenges, including variations in contrast, illumination, image quality, and the presence of artifacts, which may affect model performance and generalization. However, the use of evolutionary optimization strategies for improving ResNet-based dental caries classification remains insufficiently explored. To address this gap, the present study proposes ResNet-GA, a hybrid framework that combines the feature extraction capability of ResNet-50 with a genetic algorithm-based optimization process. This approach automatically searches for improved hyperparameter configurations and aims to enhance the accuracy, reproducibility, and robustness of automated dental caries detection.

In this study, we propose ResNet-GA, a novel optimization framework that integrates a genetic algorithm with the ResNet-50 architecture, a deep convolutional neural network widely recognized for its effectiveness in image classification and transfer learning tasks. We introduce a GA-driven optimization strategy to automatically search for improved convolutional configurations, including the number of filters and kernel sizes in selected convolutional blocks. This reduces dependence on manual and experience-based architectural tuning.

While earlier testing was conducted using a simplified 8- layer convolutional neural network to validate our evolutionary approach, all final experiments and evaluations presented in this paper were conducted exclusively on ResNet-50. The selected hyperparameters are encoded into chromosomes, evaluated through training on the dental image dataset for each individual, and evolved over multiple generations using standard genetic operations, such as selection, crossover, and mutation. This framework enables effective and scalable hyperparameter optimization for deep learning based dental diagnosis.

Although the genetic algorithm is configured with a relatively small population size and limited number of generations due to computational constraints, this design choice is consistent with recent studies in medical image analysis where full training of deep CNNs is computationally expensive. The objective is not to perform exhaustive neural architecture search, but rather to introduce a controlled and efficient evolutionary refinement of critical convolutional parameters.

Moreover, restricting the search space to filter numbers and kernel sizes ensures that the core ResNet-50 residual structure remains unchanged, preserving stability while enabling targeted optimization. Future work will extend this framework using larger populations, multi objective optimization, and distributed training strategies.

Despite extensive research in dental image analysis, a key limitation remains the absence of standardized experimental protocols, which hinders fair and direct comparison between studies. Reported performance metrics are often obtained under heterogeneous conditions, including variations in dataset composition, preprocessing strategies, and evaluation methodologies.

This lack of standardization significantly affects the reproducibility and comparability of results, highlighting the need for unified benchmarking frameworks in dental AI research.

The main contributions of this study are summarized as follows:

  1. 1.

    We propose ResNet-GA, a hybrid deep learning framework for automated dental caries detection that combines the feature extraction capability of ResNet-50 with genetic algorithm-based optimization.

  2. 2.

    We introduce a GA-driven optimization strategy that automatically searches for improved convolutional configurations, specifically the number of filters and kernel sizes in three convolutional blocks, encoded as a six-gene chromosome. Complementary hyperparameters (learning rate, batch size, dropout rate) were tuned separately through a preliminary manual search and then held fixed during the GA process, reducing the dependence on manual architectural tuning specifically.

  3. 3.

    We design a chromosome encoding scheme in which each individual represents a candidate ResNet-50 configuration. This encoding enables evolutionary search while preserving the core residual structure of ResNet-50.

  4. 4.

    We evaluate the proposed model using multiple diagnostic performance metrics, including accuracy, precision, recall, F1-score, confusion matrix, and AUC. The results are also discussed in relation to representative deep learning methods previously reported for dental caries detection.

  5. 5.

    We provide an empirical analysis showing that the integration of genetic algorithms with ResNet-50 improves classification performance and diagnostic reliability compared with non-optimized ResNet-50 configurations.

Before detailing our research topic, we clarify the terminology used throughout this article in Section II, and then review related works in Section III. Section IVdescribes our proposed model, ResNet-GA, and clarifies its steps. The experiments we conducted and the results obtained by our algorithm are presented in Section V. A summary of the lessons learned and a reflection on future research work are provided in Section VI.

Background

Building on the importance of hyperparameter tuning in improving the accuracy of neural networks, it is essential to consider the methodologies that can be employed to achieve optimal results. Among these methodologies, transfer learning and genetic algorithms have emerged as two significant approaches that can complement hyperparameter optimization efforts.

Hyperparameter tuning is a critical process in machine learning that involves adjusting the parameters that govern the training process of a model. These parameters, known as hyperparameters, include aspects such as the learning rate, batch size, and number of layers in a neural network. Proper tuning of these hyperparameters can lead to significant improvements in the model accuracy and generalization. However, finding the optimal set of hyperparameters can be challenging because of the wide search space involved. Therefore, methodologies such as transfer learning and genetic algorithms are increasingly being utilized to enhance the efficiency and effectiveness of hyperparameter optimization methods.

Transfer learning is a technique that allows researchers to leverage pretrained models, such as ResNet-50, to improve performance on new tasks. ResNet-50 is a convolutional neural network (CNN) architecture that consists of 50 layers and is designed to address the problem of vanishing gradients using its residual learning framework. This architecture has been widely used for various image recognition tasks, including object detection and image classification, owing to its ability to learn complex features from large datasets14. Transfer learning capitalizes on the knowledge gained from training a model on a large dataset and applies it to different but related tasks. ResNet-50 is a notable example of a deep learning architecture that employs residual connections to facilitate the training of extremely deep networks. Residual connections help mitigate the problem of vanishing gradients, allowing gradients to flow more easily through the network during backpropagation. This characteristic makes ResNet- 50 particularly effective for tasks that require extracting intricate features from images, such as distinguishing between different classes in image classification.

Using ResNet-50, researchers can significantly reduce the amount of labeled data required for training on new tasks, thus accelerating the learning process and enhancing the model performance. This is particularly advantageous in domains where acquiring labeled data is expensive or time-consuming, such as in medical imaging. For example, studies have shown that ResNet-50 based architectures are effectively employed for tasks such as the identification of plant diseases and medical diagnosis, demonstrating high accuracy rates even when trained with limited data15. The ability to transfer learned features from one domain to another not only speeds up the training process but also improves the robustness of the model in the target domain.

In conjunction with transfer learning, genetic algorithms provide a robust framework for hyperparameter optimization.

These algorithms are inspired by the principles of natural selection and evolution, in which potential solutions to a problem are treated as individuals in a population. Through processes such as selection, crossover, and mutation, genetic algorithms iteratively refine the population to obtain better solutions.

Genetic algorithms (GAs) are a class of optimization algorithms that mimic the process of natural evolution to solve complex problems16. In the context of hyperparameter optimization, GAs treat different hyperparameter configurations as individuals in a population. The algorithm evaluates the performance of each individual based on a fitness function, which typically measures the accuracy of the model on a validation dataset.

Through iterative processes, such as selection (choosing the best-performing individuals), crossover (combining features of two individuals to create offspring), and mutation (randomly altering some features), GAs evolve the population over successive generations. This evolutionary approach allows for a more efficient exploration of the hyperparameter space, particularly in complex models, where traditional grid search or random search methods may be computationally expensive and time-consuming. The integration of genetic algorithms with deep learning techniques can lead to improved model precision and robustness by effectively navigating complex hyperparameter landscapes17.

Together, these methodologies not only address the challenges associated with hyperparameter tuning but also contribute to the overall effectiveness of neural networks across various applications. By employing advanced techniques such as transfer learning and genetic algorithms, researchers can navigate the complexities of model training more effectively, ultimately leading to more accurate and robust machine learning solutions.

The combination of transfer learning and genetic algorithms also provides a powerful toolkit for researchers and practitioners in the field of machine learning. Transfer learning allows for the efficient reuse of knowledge from pretrained models, significantly reducing the amount of data and time required for training on new tasks. Genetic algorithms offer a systematic and adaptive approach to hyperparameter optimization, allowing the identification of optimal configurations that can enhance model performance.

As machine learning continues to evolve, integrating these methodologies will likely play a crucial role in advancing the capabilities of neural networks in diverse applications ranging from medical diagnostics to autonomous systems. By leveraging the strengths of both transfer learning and genetic algorithms, researchers can develop more sophisticated models that are not only accurate but also resilient to the challenges posed by real-world data.

Related works

In recent years, integrating transfer learning techniques for dental caries detection has garnered considerable attention in the dental health and medical imaging fields. This growing interest stems from the ability of transfer learning to leverage pretrained models on large-scale image datasets from various domains, enabling the automated and more accurate detection of dental caries. By using these models, which have already been trained on vast, diverse datasets, researchers can effectively apply them to the specific task of caries detection, significantly improving the precision and efficiency of the process. Transfer learning not only reduces the computational burden by minimizing the need for extensive data collection and model training from scratch, but also enhances the model’s ability to generalize to new, unseen data, thus improving diagnostic capabilities across diverse populations and environments.

Transfer learning studies

One notable example of this approach uses ResNet50 and MobileNet to automate dental cavity detection. The model trained with MobileNet achieved an accuracy rate of 96.12%, utilizing a dataset of 395 normal and 395 caries dental images18. While these results are promising, the relatively small dataset used for training may limit the model’s ability to generalize to a broader set of dental conditions and patient demographics. Similarly, another study employed several pretrained convolutional neural networks (CNNs), including VGG-16, VGG-19, DenseNet-121, and Inception V3, coupled with advanced preprocessing techniques such as histogram equalization and Sobel edge detection, to en- hance caries detection performance. Among these models, the VGG-16 model achieved the highest accuracy of 98.3% using the stochastic gradient descent (SGD) optimizer19. However, this study focused primarily on high-quality images, which may not reflect the variability and lower quality of real- world dental radiographs, presenting a limitation regarding model robustness for practical applications.

A different approach introduced a cost-effective virtual computer vision AI system designed to predict dental cavities from non-standardized smartphone photographs. This system utilized the You Only Look Once (YOLO) algorithm ensemble, combined with transfer learning techniques using models such as ResNet50, ResNet101, VGG16, AlexNet, and DenseNet, achieving a diagnostic accuracy of 86.96%20. While smartphone-based AI solutions offer considerable promise in improving access to dental diagnostics, particularly in resource-limited settings, the quality of images captured by consumer smartphones may vary widely, potentially impacting the reliability and accuracy of the model. Furthermore, reliance on non-standardized photographic conditions raises concerns about the consistency of results, as these images may not always meet the necessary quality standards for accurate diagnosis.

A hybrid model combining ResNet-50 for deep feature extraction and XGBoost for classification demonstrated high performance in cavity-level prediction, with an accuracy of 97.81%21. Although this hybrid approach shows promise for improving cavity detection, it also introduces complexity into the model architecture, which could lead to increased computational demands. In clinical settings, where time and resources are often limited, the use of such complex models may be challenging unless further optimizations are made to reduce the processing times and computational load. Additionally, while this model showed strong results, the dataset used in the study may not be sufficiently representative of the wide range of dental conditions encountered in everyday clinical practice.

In another study, a CNN-based classification framework was developed to identify a range of dental pathologies, including dental caries, necrotic teeth, gingival inflammation, and herpes simplex. This system achieved perfect accuracy, recall, and F1-scores for detecting dental caries, gingival inflammation, and healthy dentition22. However, the system’s ability to generalize to more complex or rare conditions remains unclear because the study primarily focused on more common dental diseases. Additionally, the dataset used may not have captured the full diversity of cases seen in clinical environments, raising questions about the robustness of the model when applied to a more varied patient population.

Additionally, Faster R-CNN models with ResNet-50 and ResNet-101 configurations were applied to detect and localize dental caries, with ResNet-50 showing superior performance regarding the mean average precision23. While Faster R-CNN is known for its efficiency in object detection tasks, the localization of dental caries within an image remains challenging, particularly in cases where the caries are subtle or in the early stages. The model’s performance could be affected by image quality, the presence of overlapping structures, or the degree of caries progression, highlighting the need for further refinement and testing in clinical settings.

Another approach involved modifying DenseNet-121 and applying transfer learning to classify oral pathologies, such as cavities, gingivitis, and calculus, achieving an impressive accuracy of 95% on a limited dataset of 992 images24. Although DenseNet-121 performed well, the study’s reliance on a limited dataset raises concerns regarding the model’s generalizability. More extensive and diverse datasets are necessary to ensure that the model can effectively classify oral pathologies across various populations and clinical scenarios. Finally, a hybrid model combining segmentation and detection techniques using a ViT and U-Net demonstrated an extraordinary accuracy rate of 99%, showcasing the high potential of AI to enhance the efficiency and accuracy of dental diagnostics25. Although the high accuracy of this model is promising, the increased complexity associated with combining multiple deep learning models (ViT and U-Net) could result in increased computational requirements and longer processing times, which may not be ideal for real-time clinical applications. Furthermore, the model’s performance on a wider range of dental conditions and image qualities should be evaluated to confirm its practical utility in diverse diagnostic scenarios.

Recent studies have further extended transfer-learning approaches to panoramic radiographs and to acquisition modalities beyond the standard bitewing or periapical format used in most of the studies above. The authors of26 developed a modified DenseNet-121-based framework with a configured classifier for caries detection from panoramic images and, using a large dataset of cropped tooth images, achieved improved classification performance compared with conventional CNN architectures; however, the classifier structure was manually designed, leaving automatic, evolutionary optimization of the network configuration unexplored a gap directly addressed by the GA-based hyperparameter search proposed in the present study. The researchers in27 introduced a large-scale benchmark for pediatric dental caries detection on panoramic radiographs (the MD-OPG dataset),

Table 1 Some deep learning models used for the classification of dental caries in the literature.

Full size table

systematically comparing several deep learning architectures and highlighting challenges specific to pediatric populations, such as smaller occlusal surfaces, mixed dentition, and greater inter-patient variability. The study of28 proposed a Transformer based detection model targeting early-stage caries in panoramic radiographs, reporting improved sensitivity for subtle, low contrast lesions relative to conventional CNN baselines suggesting that Transformer architectures may offer complementary benefits for early-lesion detection specifically, even though CNN backbones remain broadly competitive for lesion-level tasks overall29.

A related but distinct strategy for addressing the limited size of the dental imaging datasets used in the studies above is generative data augmentation. The researchers in30 proposed Pulp-Caries-GAN, a generative adversarial network integrating a biomimetic metaheuristic optimization strategy modeled on the spatial coupling dynamics of dental pulp tissue to generate synthetic dental caries images and improve downstream segmentation performance, raising the Dice coefficient from 89.75% to 95.12% on a limited annotated pediatric dataset of only 193 images, with 87% of the synthetic images judged clinically indistinguishable from real radiographs by expert dentists.

Beyond radiographic imaging, several recent studies have explored caries detection using alternative acquisition modalities. The research of31 evaluated a YOLOv8x based model for detecting new clinical caries directly from intraoral photographs of school-aged children, reporting encouraging field-deployable performance despite the variability inherent to non-standardized photographic conditions. The authors of32 similarly proposed an improved YOLOv8-based intelligent detection system for broader dental health status assessment, while the research in33 applied YOLOv9 to X-ray image analysis for automated dental diagnosis, further confirming the continued and expanding adoption of YOLO family object detectors across diverse dental imaging tasks. Moving beyond conventional two dimensional imaging altogether, the study by34 combined spectrophotometric CIELAB colour measurements with machine learning classifiers to grade occlusal caries severity, and the work of35 developed an automatic detection and quantitative analysis method for early stage caries using polarization sensitive optical coherence tomography (PS-OCT). Together, these studies illustrate that optical, colorimetric, and non-ionizing modalities can provide complementary alternatives to conventional radiographic detection, particularly for very early or subtle lesions, though like the panoramic based approaches above, they continue to rely on fixed architectures or manually configured classifiers rather than automated hyperparameter optimization.

Table 1 summarizes a selection of previous studies that explored the application of deep learning methodologies in dental diagnostics, particularly focusing on the classification and detection of dental caries in dental images. Each study employs different methodologies, datasets, and evaluation metrics, reflecting the diversity of approaches in this rapidly evolving field.

Recent advanced architectures for dental image analysis

Recent studies have also explored more advanced architectures for dental image analysis, including Transformer-based, hybrid CNN–Transformer, and ensemble learning models. The authors of29 compared CNNs, Transformer-based architectures, and hybrid models for dental image segmentation tasks, including tooth segmentation, tooth-structure segmentation, and caries-lesion segmentation. Their findings showed that CNN-based architectures remained highly competitive and even outperformed Transformer-based and hybrid models in several dental segmentation tasks, suggesting that model complexity does not always guarantee superior performance in dental imaging.

Recent systematic evidence has also confirmed the rapid development of deep learning methods for caries diagnosis using bitewing radiographs. The researchers in39 reviewed recent deep learning approaches for detecting, segmenting, and classifying dental caries on bitewing radiographs, showing that architectures such as ResNet and YOLO are among the most frequently used models in this field. These findings highlight the growing interest in automated caries diagnosis, but they also indicate that model performance remains strongly dependent on dataset size, annotation quality, imaging modality, and evaluation protocol.

In addition, CNN-based models remain relevant for bitewing image analysis. The authors of40 proposed a convolutionalneural network based approach for dental caries diagnosis from bitewing images, further supporting the suitability of CNN architectures for radiographic caries classification. However, despite these recent advances, many existing studies still rely on fixed architectures or manually selected hyperparameters. In contrast, the present study adopts a ResNet-50 backbone and improves its performance through genetic algorithm based hyperparameter optimization. This strategy aims to preserve the feature extraction capacity of ResNet-50 while reducing dependence on empirical and experience-based hyperparameter selection.

Discussion of dental image analysis studies

The reviewed studies demonstrate the growing importance of deep learning methods in dental image analysis, particularly for the detection, classification, and localization of dental caries. Earlier studies mainly relied on CNN-based and transfer learning architectures, either independently or in combination with other machine learning techniques. For instance, the study referenced as18 used ResNet50 and MobileNet for dental caries detection and achieved an accuracy of 96.12% using a dataset composed of 395 normal and 395 caries dental images. Similarly, the use of pretrained models such as VGG-16, VGG-19, DenseNet-121, and Inception V3 in19 achieved an accuracy of 98.3%, confirming the effectiveness of transfer learning for extracting discriminative features from dental images.

Other studies have extended CNN-based approaches by integrating object detection or hybrid classification methods. For example, the work of20 combined YOLO-based models with transfer learning architectures, including ResNet50, VGG16, AlexNet, and DenseNet, and achieved a diagnostic accuracy of 86.96% using non-standardized smartphone photographs. This result highlights the potential of deep learning for accessible and low-cost dental screening. However, it also shows that image acquisition conditions, illumination, resolution, and standardization can strongly affect diagnostic performance. Similarly, the combination of ResNet-50 for deep feature extraction and XGBoost for classification in21 achieved an accuracy of 97.81%, showing that hybrid learning strategies can improve classification performance, although they may also increase computational complexity.

The study in22 reported very high performance for detecting dental caries and other oral pathologies using a CNN-based classification framework. In addition, Faster R-CNN models combined with ResNet-50 and ResNet-101 were used in23 to detect and localize dental caries, demonstrating the usefulness of object detection frameworks in dental imaging. DenseNet-121 was also applied in24 for classifying oral conditions, achieving 95% accuracy.

Furthermore, the ViT and U-Net-based hybrid model in25 showed the potential of combining segmentation and detection strategies for dental diagnosis. These studies collectively indicate that CNNs, transfer learning models, object detection frameworks, and hybrid architectures can all contribute to automated dental caries diagnosis.

The recent literature further confirms that more advanced architectures, including Transformers and hybrid CNN–Transformer models, are increasingly being investigated in dental image analysis. However, recent comparative evidence suggests that more complex architectures do not necessarily guarantee superior performance. For example, the authors of29 compared CNNs, Transformer based models, and hybrid architectures for dental image segmentation tasks and showed that CNN-based architectures remained highly competitive. This finding is important for dental caries detection because carious lesions are often small, low-contrast, and spatially localized; therefore, strong local feature extraction remains essential.

Recent systematic evidence also supports the relevance of deep learning for caries diagnosis on bitewing radiographs. The researchers of39 reviewed recent deep learning approaches for detecting, segmenting, and classifying caries lesions in bitewing radiographs. Their review shows that deep learning has become a promising tool for radiographic caries diagnosis, but it also highlights persistent challenges related to dataset size, annotation quality, imaging variability, class imbalance, and differences in evaluation protocols. These issues limit the direct comparability of reported results across studies and underline the need for robust and reproducible model development strategies.

In addition, The authors of40 demonstrated the usefulness of CNN-based models for dental caries diagnosis from bitewing images, further supporting the suitability of convolutional architectures for this task. Taken together, the recent studies29-40 indicate that CNN based methods remain highly relevant in dental imaging, even in the presence of more complex Transformer based and hybrid models. This supports the choice of ResNet-50 as the backbone architecture in the present study.

While the studies reviewed above report high accuracy on curated, single source datasets, recent clinical evaluations highlight a substantial gap between benchmark performance and real-world diagnostic agreement. The researchers in36 evaluated an AI-assisted caries detection system (Diagnocat) on full arch CBCT examinations by comparing its outputs with clinical assessments performed by three dentists on 100 patients, reporting an overall diagnostic accuracy of 89%, with a sensitivity of only 50.8% despite a high specificity of 97.2%, and showing that performance varied according to tooth type, patient age, gender, and the examining clinician. A related systematic review and meta-analysis by the study of37 synthesized evidence from 17 studies investigating deep learning based caries detection using Faster R-CNN and Mask R-CNN, demonstrating that Mask R-CNN achieved superior pooled sensitivity, specificity, and AUC compared with Faster R-CNN, while emphasizing substantial heterogeneity among studies due to variations in imaging modalities, annotation protocols, dataset characteristics, and validation strategies.

Reinforcing this concern, a systematic review by the authors of38 assessed the translational readiness of imaging-based artificial intelligence systems across dentistry, concluding that despite consistently strong algorithmic performance reported in the literature, most studies fall short of demonstrating genuine readiness for routine clinical integration, owing to limited external validation, heterogeneous reporting standards, and a scarcity of prospective clinical trials. In a closely related direction, the authors of41 directly compared human examiner and AI performance in radiographic caries detection against ex vivo tooth-section histological ground truth, showing that AI systems can match or even exceed human sensitivity for certain lesion categories while still exhibiting systematic discrepancies relative to the histological reference standard. Together, these findings reinforce that robust dental AI models require validation under realistic clinical conditions rather than relying exclusively on controlled benchmark datasets a consideration directly relevant to the present study, which relies on a single, curated Kaggle dataset.

Overall, the reviewed literature shows that deep learning models have achieved promising results in dental caries detection and classification. Nevertheless, several limitations remain. Many studies depend on fixed architectures, manually selected hyperparameters, limited datasets, or computationally demanding hybrid and ensemble models. Moreover, the comparison between studies is often difficult because of differences in datasets, image modalities, preprocessing strategies, annotation procedures, and evaluation metrics. In this context, the proposed ResNet-GA framework addresses an important gap by preserving the strong feature extraction capability of ResNet-50 while using a genetic algorithm to automatically optimize key hyperparameters. This strategy aims to improve classification performance, reduce dependence on manual tuning, and enhance the reproducibility of automated dental caries detection.

Hyperparameter optimization studies

In addition, optimizing the hyperparameters of neural networks is a significant area of research that aims to improve model performance and accuracy. Various methodologies have been proposed to enhance the efficiency and effective- ness of the hyperparameter tuning. For example, Gaussian processes were utilized for hyperparameter optimization in dialog act classification, leading to notable improvements in F1 scores in multiple datasets and demonstrating the efficacy of probabilistic models in this context42. Similarly, the HOAX package was introduced to support hyperparameter optimization for various neural network architectures, pro- viding researchers with a flexible tool to explore different configurations43.

Another notable approach was presented by1, who explored methods to accelerate hyperparameter optimization, specifically for deep convolutional neural networks (CNNs). Their research highlighted the applicability of various optimization algorithms, including random search and genetic algorithms, to enhance the tuning processes. Further contributions to this field emphasized the role of performance prediction in the search for neural architecture, with the aim of reducing human intervention by automating the hyperparameter optimization process15. This aligns with the findings of44, who proposed a genetic algorithm-based optimization approach for CNNs, illustrating the potential of evolutionary strategies to navigate the hyperparameter space.

Recent studies have also confirmed the relevance of metaheuristic and evolutionary optimization strategies for improving deep learning models. The authors of45 proposed an improved salp swarm algorithm-driven deep CNN for brain tumor analysis, showing that swarm-based optimization can be effectively used for CNN hyperparameter tuning in medical imaging. Although their work was applied to brain MRI rather than dental radiographs, it supports the broader relevance of biologically inspired optimization for medical-image classification.

Similarly, recent optimization research has explored particle swarm optimization and evolutionary neural architecture search as alternatives to manual model design. The researchers of46 introduced a quantum-inspired adaptive mutation operator enabled PSO for parameter optimization, while the study of47 proposed an efficient evolutionary neural architecture search method designed to reduce the computational burden of architecture evaluation. These studies show that evolutionary and swarm based approaches are increasingly used to improve neural network design and parameter selection. However, most of these works focus either on general optimization, architecture topology search, or non-dental medical datasets. In contrast, the proposed ResNet-GA framework preserves the core ResNet-50 architecture and applies a genetic algorithm to optimize selected hyperparameters specifically for dental caries classification.

The authors of48also focused on optimizing hyperpa- rameters in CNNs, underscoring the complexity of finding an optimal set of parameters that produce accurate models within reasonable timeframes. They introduced an improved genetic algorithm that optimizes hyperparameters based on a fractal loss function, further illustrating the various strategies employed to improve model accuracy. Furthermore49, applied a goal oriented sensitivity analysis to hyperparameter optimization, revealing the intricate relationships between hyperparameters and model performance.

The exploration of bilevel optimization, as discussed by Mackay, presents another innovative approach to hyperparameter tuning, where the optimization of hyperparameters is treated as a hierarchical problem, allowing for more nuanced adjustments based on the model performance50. This is complemented by the work of51, who demonstrated the effectiveness of Bayesian approaches in optimizing hyperparameters for neuromorphic hardware and achieved significant accuracy improvements.

However, several challenges persist in these hyperparameter optimization techniques. The computational demands associated with tuning can be substantial, often requiring extensive resources and time that may not be feasible for all users to afford. In addition, many methods are based on heuristic or empirical strategies that lack a strong theoretical foundation, which can lead to inconsistent results across applications or datasets. Furthermore, while automation seeks to streamline processes, it may unintentionally overlook critical insights from human expertise that could refine tuning efforts. Lastly, there remains notable opacity regarding how specific hyperparameter choices impact overall model performance, complicating efforts to replicate findings or fully understand the underlying dynamics.

Discussion of hyperparameter oprimization studies

Table 2 summarizes the main optimization strategies used to improve the performance of neural networks, including probabilistic methods, genetic algorithms, sensitivity analysis, bilevel optimization, Bayesian optimization, swarm intelligence, and evolutionary neural architecture search. These studies collectively show that hyperparameter optimization has become a central issue in deep learning because model performance is highly sensitive to parameters such as learning rate, batch size, dropout rate, kernel size, number of filters, and architecture-related settings.

Table 2 Summary of methods for hyperparameter optimization.

Full size table

The studies referenced as42 and43 illustrate the usefulness of probabilistic and software-assisted optimization approaches. Gaussian processes were used in42 to model uncertainty during hyperparameter selection and improve F1 scores in dialog act classification, whereas the HOAX package in43 provided a flexible tool for exploring hyperparameter configurations across neural network architectures. These methods show that systematic hyperparameter search can reduce manual intervention and improve model performance.

Other works have focused on random search, genetic algorithms, and sensitivity based approaches. For example1, investigated methods for accelerating hyperparameter optimization in deep convolutional neural networks, including random search and genetic algorithms. Similarly44, and48 showed that genetic algorithm based strategies can be effective for CNN hyperparameter tuning. Goal oriented sensitivity analysis49 further highlighted the complex relationship between hyperparameters and model performance, while bilevel optimization50 and Bayesian optimization51 provided more structured frameworks for optimizing hyperparameters in complex neural models.

The newly added studies further confirm the increasing relevance of metaheuristic and evolutionary optimization in deep learning. The authors of45 proposed an improved salp swarm algorithm-driven deep CNN for brain tumor analysis, showing that swarm-based optimization can improve CNN hyperparameter tuning in medical imaging. Although their work focused on brain MRI rather than dental radiographs, it is relevant because it demonstrates the effectiveness of biologically inspired optimization strategies for medical image classification. The study in46 introduced a quantum inspired adaptive mutation operator enabled PSO for optimizing parameters of Kolmogorov–Arnold networks, further illustrating the recent development of adaptive and swarm-based optimization methods. In addition, the authors of47 proposed an efficient evolutionary neural architecture search algorithm without full training, aiming to reduce the computational cost usually associated with architecture evaluation.

These recent works demonstrate that optimization research is moving toward more adaptive, evolutionary, and computation-aware strategies. However, several limitations remain. Some approaches are applied to non dental medical datasets, such as brain MRI; others focus on general neural architectures or parameter optimization rather than dental caries classification. In addition, neural architecture search methods often optimize the global network topology, which may require high computational resources or produce architectures that are difficult to interpret and reproduce. Therefore, although these studies confirm the value of evolutionary and metaheuristic optimization, they do not directly address the specific problem targeted in this work.

In contrast, the proposed ResNet-GA framework adopts a more focused optimization strategy. Rather than searching for an entirely new architecture, it preserves the core residual structure of ResNet-50 and uses a genetic algorithm to optimize selected hyperparameters that directly influence dental caries classification performance. This approach provides a balance between model performance, architectural stability, and computational feasibility. It also reduces dependence on manual hyperparameter selection and improves the reproducibility of the model development process.

Overall, the reviewed optimization studies support the relevance of using genetic algorithms for deep learning enhancement. Nevertheless, the specific integration of genetic algorithm based hyperparameter optimization with ResNet-50 for dental caries detection remains insufficiently explored. This gap justifies the proposed ResNet-GA framework and strengthens its contribution to automated dental image analysis.

Recent studies (39,40,41,42,43,44,–45) have further confirmed the relevance of deep-learning architectures for dental image classification. These works reported promising results using convolutional neural networks, transfer learning models, and hybrid deep learning approaches. However, direct comparison between these studies remains difficult because of differences in dataset size, image acquisition conditions, preprocessing procedures, class distribution, and evaluation protocols. In addition, many existing approaches rely on fixed architectures or manually selected model configurations, which may limit their adaptability to specific datasets. In this context, the proposed ResNet-GA framework complements previous contributions by introducing a GA-based optimization process aimed at improving model configuration and classification performance.

Proposed RESNET-GA algorithm

Our research is notable for its focus on ResNet-GA, a deep learning based system for diagnosing dental caries, and the novel exploration of genetic algorithms for the optimization and performance enhancement of the model. The significance of using deep learning models, particularly ResNet architectures, for medical imaging has been well documented. For example, studies have shown that deep learning techniques can significantly improve the precision of dental disease detection, including dental caries, by analyzing complex patterns in imaging data. The research process involved several key steps, as illustrated in Fig. 1: (i) Dataset Preparation, (ii) Image Preprocessing (images were normalized and resized to ensure consistency and compatibility with the ResNet-50 architecture), (iii) ResNet-50 Initialization and Pre-training (the ResNet-50 model was initialized and pre-trained on the prepared training set to provide a starting point for further optimization), (iv) hyperparameter Optimization with Genetic Algorithms (we employed a genetic algorithm to optimize the hyperparameters of the ResNet-50 model), (v) Evaluation and Selection, (vi) Crossover and Mutation, and (vii) New Generation and Stopping Criteria.

The initial phase of our methodology involved the preparation of a dataset, which is crucial for training effective deep learning models. The quality and size of the dataset directly influence the model performance, as demonstrated by previous research that emphasizes the importance of comprehensive datasets in achieving high diagnostic accuracy52. Subsequently, image preprocessing was performed to ensure that the input data were suitable for the ResNet-50 architecture, which is known for its ability to handle large scale image classification tasks53.

In the initialization and pretraining phases, the ResNet-50 model was trained on the prepared dataset. This step is essential because it allows the model to learn the initial features from the data, which can then be fine-tuned through further training54. The subsequent optimization of hyperparameters using genetic algorithms represents a novel approach in this study. Genetic algorithms are particularly effective in exploring large search spaces and have been successfully applied to optimize machine learning models in various domains, including healthcare55.

Initially, five ResNet-50 models were generated using transfer learning from our main neural network. These models served as the initial population for the genetic algorithms. By applying selection, crossover, and mutation, we aimed to generate ResNet models with improved accuracy and recognition rates for diagnosing dental caries. This evolutionary approach not only improves the performance of the model but also contributes to the robustness of the diagnostic system, making it a valuable tool in clinical settings53. Our research integrates advanced deep learning techniques with genetic algorithms to enhance the diagnostic capabilities of dental caries, addressing a critical need for the management of dental health. Combining these methodologies improves diagnostic accuracy and efficiency, ultimately benefiting patient care.

The pseudocode provided in ResNet-GA outlines a methodology for discovering an optimal ResNet-50 architecture designed for a specific task using a genetic algorithm. This algorithm treats ResNet-50 architectures as individuals within a population that are subject to evolutionary principles such as selection, crossover, and mutation. The fitness of each individual is evaluated based on its performance on the given task, guiding the selection process to favor architectures with superior performance. The iterative refinement of the population through successive generations ultimately converges to a highly optimized ResNet-50 architecture that excels in the desired application.

Algorithm 1

ResNet-GA: hyperparameter optimization using genetic algorithm.

Fig. 1

Proposed architecture of ResNet-GA.

Data collection

As illustrated in Fig. 2, the dataset used in this study consists of 1554 color images of dental caries, collected from several datasets on the Kaggle platform. The dataset was divided into two parts: a training set containing 1260 images (945 depicting caries and 315 depicting non-caries) and a testing set comprising 294 images (210 depicting caries and 84 depicting noncaries). This distribution is crucial to ensure that the model is adequately trained to recognize both classes, which is essential for achieving high diagnostic accuracy.

The dataset used was well-structured and sourced from a reputable platform, providing a solid foundation for training and evaluating deep learning models aimed at diagnosing dental caries. Careful consideration of dataset composition, quality, and accessibility underscores the importance of data in developing effective machine learning solutions in dentistry.

Fig. 2

Used teeth decay dataset (caries and non-caries).

Preprocessing

We used data preprocessing techniques to enhance the performance of our model. This involved two essential methods: image resizing and normalization. Because our dataset contained images with varying widths and heights, resizing them to a uniform size was crucial for optimal performance. This ensured computational efficiency, prevented distortions, and facilitated better feature extraction by the convolutional layers of the ResNet-50 model. In this study, we used the target size function to resize all color images to \(224 \times 224\) pixels, which is the standard input size for ResNet-50. This specific dimension is widely recognized in the deep learning community, as it aligns with architectural design, allowing the model to process and learn efficiently from input data.

Additionally, we applied a standardization technique to normalize the pixel values, ensuring that they fell within a particular range, typically between 0 and 1 or \(-1\) and 1. Normalization is a critical preprocessing step that enhances the stability of the model training and convergence speed. By scaling the pixel values, we mitigated the impact of varying lighting conditions and color intensities present in the dataset, which can otherwise lead to suboptimal model performance. Normalization also helps prevent issues related to gradient descent, where large input values can lead to explosive gradients, thereby hindering the learning process55.

The combination of resizing and normalization not only improves the computational efficiency of the model but also enhances its ability to generalize to the unseen data. Preprocessing steps, such as image normalization or resizing, are commonly employed in deep learning pipelines for medical imaging, as illustrated in52. By ensuring that the input data are consistent and well-scaled, we created a more robust foundation for the ResNet-50 model to learn effectively from the training dataset.

Resnet50 parameters

Hyperparameter tuning is crucial for optimizing the performance of deep learning models, such as ResNet-50. Adjusting various hyperparameters, such as the learning rate, batch size, and number of filters, can significantly affect the accuracy and generalizability of the model56. However, finding the optimal set of hyperparameters is complex and often relies on experience and experimentation rather than theoretical knowledge57. In our study, we used a genetic algorithm to automate the exploration of convolutional configurations. Specifically, the GA searched over the number of filters and kernel sizes in three convolutional blocks, encoded as a six-gene chromosome. The learning rate, batch size, and dropout rate were determined beforehand through a separate manual tuning stage and then held fixed at the values reported in Table 3\((learning rate = 0.001, batch size = 32, dropout rate = {0.5}\) for all GA individuals, so that the evolutionary search focused exclusively on the convolutional architecture. The final output layer uses a Sigmoid activation function to perform binary classification between caries and non-caries cases. During each generation of the GA, individuals representing different ResNet configurations (with distinct combinations of these hyperparameters) were trained for 20 epochs, and their validation accuracy was used as the fitness value. The GA then applies selection, crossover, and mutation operators to evolve the population toward more optimal hyperparameter combinations58.

Table 3 presents a selection of different parameter sets obtained by ResNet-50 through the genetic algorithm based hyperparameter tuning process. The use of a genetic algorithm for hyperparameter optimization is particularly advantageous because it can efficiently navigate the vast search space associated with deep learning models, reducing the time and computational resources required compared with traditional methods such as grid search or random search59.

Table 3 Example of optimized hyperparameters and layer configuration by genetic algorithm.

Full size table

It should be clarified that the genetic algorithm does not modify the internal architecture of the pretrained ResNet-50 backbone. Instead, a pretrained ResNet-50 model initialized with ImageNet weights is employed as a fixed feature extractor to preserve the rich visual representations learned through transfer learning. The genetic algorithm operates exclusively on the task-specific layers appended after the ResNet-50 feature extractor, including the convolutional and classification layers. Since these newly introduced layers have no corresponding pretrained parameters, they are initialized using He normal initialization before training. Consequently, all candidate networks share the same pretrained ResNet-50 backbone while differing only in the architectural configuration of the GA-optimized layers. This design preserves the benefits of transfer learning while enabling evolutionary optimization without introducing dimensional incompatibilities between pretrained and newly created layers.

Chromosome encoding and population initialization

In the initialization stage, the genetic algorithm chromosomes encode the architectural configuration of the task-specific convolutional layers appended after the pretrained ResNet-50 backbone. The pretrained ResNet-50 served as the backbone network, providing robust feature representations learned from ImageNet. Instead of modifying the pretrained backbone, the genetic algorithm was designed to optimize only the task-specific layers appended after the ResNet-50 feature extractor. Although random weight initialization is an option, the use of transfer learning offers several advantages, particularly when working with limited datasets. By adopting the knowledge encoded in the pretrained model’s weights, we can achieve faster training convergence, improve generalization, and perform better than training from scratch. This approach mirrors human problem-solving strategies, where we often draw upon existing knowledge and tools to tackle new challenges efficiently.

In the proposed framework, all candidate networks share the same pretrained ResNet-50 backbone initialized with ImageNet weights. Since the convolutional and classification layers introduced by the genetic algorithm do not have corresponding pretrained parameters, these layers are initialized using He normal initialization before training. Consequently, the transferred knowledge is preserved within the ResNet-50 backbone, whereas the newly introduced task-specific layers are optimized through the evolutionary search. It is important to note that the chromosome encodes only the architectural parameters of these task-specific convolutional layers appended after the pretrained ResNet-50 backbone; the pretrained backbone itself remains unchanged and preserves its transferred ImageNet weights throughout the evolutionary optimization process. Preserving this pretrained backbone is particularly beneficial in medical imaging applications such as dental caries screening, where acquiring large labeled datasets is often challenging and expensive60.

In the proposed GA-based optimization process, each individual represents one candidate convolutional configuration encoded as a six-gene chromosome:

\(Chromosome = [F1, F2, F3, K1, K2, K3]\)

where F1, F2, and F3 denote the number of filters in the first, second, and third convolutional blocks, respectively, while K1, K2, and K3 denote their corresponding kernel sizes. The number of filters was initialized within the range 1–99, and the kernel sizes were initialized within the range 1–19. The kernel-size search space was intentionally defined over a broad interval to maximize the exploratory capability of the genetic algorithm rather than to favour large convolutional kernels. A broader search space enables the evolutionary process to investigate receptive fields of different spatial scales without introducing prior architectural bias. Importantly, allowing kernel sizes up to \(19 \times 19\) does not imply that large kernels are systematically selected or preferred: candidate solutions are retained exclusively according to their fitness values, and any configuration employing larger kernels survives only if it provides a measurable improvement in validation performance. Consequently, the genetic algorithm naturally discards configurations that do not contribute to better classification performance, regardless of kernel size. This behaviour is confirmed by the chromosomes actually selected in the five independent runs reported in Table 9: none of the runs converged to the maximum available kernel size in all three blocks, and the best-performing run (seed 3) retained kernel sizes of 1, 9, and 13. For illustration purposes, the chromosome [68, 68, 10, 15, 18, 6], which is used in the following description of the genetic operators, represents a configuration with 68, 68, and 10 filters in the three convolutional blocks, using kernel sizes of \(15 \times 15\), \(18 \times 18\), and \(6 \times 6\), respectively.

Unlike traditional hyperparameter optimization approaches such as grid search or Bayesian optimization, the proposed genetic algorithm operates on a structured chromosome encoding that preserves the architectural integrity of ResNet-50. This design constrains the search space to meaningful architectural variations, ensuring that optimization does not result in arbitrary or unstable model modifications. As a result, the proposed approach improves both interpretability and stability while maintaining the core residual learning structure. To improve the clarity and reproducibility of the genetic algorithm implementation, Table 4 presents the chromosome encoding scheme used in this study. Each chromosome consists of six genes representing the number of filters and kernel sizes of the three convolutional blocks optimized by the genetic algorithm.

Table 4 Chromosome encoding scheme used in the genetic algorithm.

Full size table

The initial population consisted of five randomly generated individuals. Each individual was trained for 20 epochs, and the classification accuracy obtained on a held-out validation split (derived fromthe training data, never from the independent test set) was used as the fitness value. The individuals were then ranked according to their validation-fitness scores, and the best-performing configurations were selected as parents for the next evolutionary step.

Individual evaluation (Fitness function)

To evaluate the fitness of each chromosome, we used a learning strategy that involved training the corresponding child network for 20 epochs using an Adam optimizer with a step decay learning rate. The performance of the child network was assessed on a held-out validation subset, obtained by splitting the training data (with stratification to preserve class proportions); the independent 294-image test set was never used during the genetic search. Specifically, the classification accuracy achieved on this validation subset determined the chromosome’s fitness value, reflecting how well the corresponding architecture generalizes beyond the data it was trained on rather than merely how well it fits the training data. The ranking process, based on these validation-fitness values, ensures a fair selection within the GA. Individuals with higher validation accuracy (better fitness) are favored for reproduction, generating subsequent generations of individuals. This selection mechanism is crucial because it drives the evolutionary process, allowing the genetic algorithm to converge towards architectures that generalize well to unseen dental images, rather than architectures that merely memorize the training set21.

Once the genetic search converged, the selected best-performing chromosome for each independent run was retrained and evaluated exactly once on the untouched, independent 294-image test set. This final evaluation involved calculating the test accuracy and constructing a confusion matrix to analyze classification performance, providing insight into the model’s ability to distinguish between caries and non-caries cases in terms of true positives, false positives, true negatives, and false negatives55.

The evaluation of the fitness of each chromosome through training and validation-based performance metrics is a critical component of our genetic algorithm approach. By separating architecture selection (validation-based) from final reporting (test-based), we ensured that the evolutionary process identified architectures based on their ability to generalize, and that the independent test set remained untouched until the single, final evaluation of the winning architecture of each run.

Selection

The initial population of the genetic algorithm consisted of five randomly generated individuals \((N = 5)\), each representing a unique convolutional configuration, consistent with Algorithm 1. We trained each individual using the training data and evaluated its performance by calculating its accuracy on a held-out validation split derived from the training data (stratified to preserve class proportions); the independent 294-image test set was not used at this stage. This validation accuracy served as the fitness value for each chromosome, providing a quantitative measure of the generalization performance of each architecture on the dental caries classification task.

Following the evaluation, we used a selection process to identify the top performers for the reproduction. In descending order, we sorted the chromosomes according to their fitness values (validation accuracy) and selected the top n chromosomes with the highest fitness values. These chosen individuals, known as parent networks, were then used in the subsequent crossover and mutation steps to create offspring for the next generation. This strategy, known as truncation selection, ensured that the genetic algorithm prioritized architectures with better validation performance, guiding the population toward improved generalization in the classification of dental caries images.

Truncation selection is a widely used method in genetic algorithms because it simplifies the selection process by focusing on a subset of the population demonstrating superior performance. This approach not only enhances the likelihood of producing high quality offspring but also accelerates the convergence of the algorithm towards optimal solutions. By emphasizing the best performing individuals, we can effectively explore the architecture space while maintaining diversity within the population, which is crucial for avoiding a premature convergence.

To assess the stability and reproducibility of the proposed framework, the complete procedure described above (population initialization, validation-based fitness evaluation, selection, crossover, and mutation) was independently repeated five times using different random seeds. Each of the five independent runs converges to its own best performing chromosome; the winning chromosome of each run is then retrained and evaluated exactly once on the untouched test set, and the resulting five test accuracies are summarized in Table 9.

Crossover

The crossover operation is crucial for generating offspring with diverse genetic information. Our study used a single point crossover61, where two parent chromosomes (selected during the previous step) exchanged genetic material at a specific, predetermined point. This process results in two child networks that inherit traits from both parents, potentially leading to novel architectures with improved performance.

Although a randomly selected crossover point is a common approach, we opted for a fixed one because of the specific structure of the chromosome encoding and considerations related to the ResNet-50 architecture. This decision was informed by the need to maintain structural integrity and ensure that the critical components of the architecture were preserved during the transition process. By introducing genetic diversity into the population, crossover facilitates the exploration of a more comprehensive range of potential solutions within the search space, increasing the chances of discovering highly optimized ResNet-50 architectures for dental caries classification62.

The use of single point crossover allows for a straight forward and effective means of combining the strengths of two parent architectures. This method can lead to offspring that possess advantageous traits from both parents, thereby enhancing the overall performance of the genetic algorithm63. Moreover, the fixed crossover point can help mitigate the risk of producing offspring that are too dissimilar from their parents, which can occur with random crossover points64.

In the context of neural network optimization, crossover operations have been shown to improve the exploration capabilities of genetic algorithms, enabling them to navigate complex architecture spaces more effectively. By generating diverse offspring, the crossover process contributes to maintaining genetic diversity within the population, which is essential to avoid premature convergence and ensure that the genetic algorithm can continue to explore new and potentially superior architectures65.

Mutation

We used a mutation operation to enhance the exploration of the search space and introduce diversity. In our implementation, we used multipoint mutation, which randomly alters the values within a chromosome. Specifically, for each individual, we randomly selected a layer and modified its associated hyperparameter (filter size or number of filters) by adding or subtracting a small randomly generated value. This approach allows for a more nuanced exploration of the architecture space, which can lead to significant variations in the offspring that may improve performance in the dental caries classification task.

Mutation is a critical component of genetic algorithms, as it prevents the algorithm from converging prematurely to a suboptimal solution. By introducing new genetic variations into the population, mutations ensure continuous exploration of the search space and increase the likelihood of discovering novel architectures with superior performance in the classification of dental caries. The introduction of small perturbations in the hyperparameters can help the algorithm escape local optima, thereby enhancing its ability to find globally optimal solutions.

In addition, the use of multipoint mutations allows multiple changes within a single individual, which can significantly increase the genetic diversity within the population. This diversity is essential for maintaining a robust search process because it enables the genetic algorithm to explore a wider range of potential solutions. Research has shown that effective mutation strategies can improve performance in various optimization tasks, including those involving neural networks.

New generation

The parent and child populations were combined to form a new generation following the crossing and mutation steps of the algorithm. The new population was then reduced to its original size by selecting individuals with the highest fitness value. The fitness of all individuals within this new population was evaluated using the same training and testing processes for the dental caries classification task.

The entire process of selection, crossover, mutation, and evaluation was repeated iteratively, with each iteration bringing us closer to the optimal solution. This continued until one or more of the stopping criteria were met. These criteria included scenarios such as the maximum fitness value of the population reaching a stable plateau, indicating convergence towards an optimal solution, or reaching a predefined maximum number of generations or time limits.

This iterative approach is fundamental to the effective- ness of genetic algorithms, as it allows for the continuous refinement of the population. By systematically evaluating and selecting the best performers, we ensured that the algorithm maintained a focus on high quality solutions while exploring the search space for potential improvements. The convergence criteria are particularly important because they help prevent unnecessary computations and ensure that the algorithm ends when satisfactory solutions are identified.

Stopping criterion

The genetic search was terminated after the predefined maximum of two generations ((G = 2)). Across the five independent runs, the best validation fitness achieved during the search ranged from 90.87% to 94.44%, with a mean of 93.02% (Table 9). Given the deliberately limited number of generations, these results should not be interpreted as evidence of full convergence of the genetic algorithm. Accordingly, a generation-by-generation convergence analysis is not presented. Instead, Table 9 and Fig. 3 report the best validation fitness obtained within the predefined search budget together with the independent test accuracy of the corresponding final retrained chromosome for each of the five runs. This multi-run evaluation provides a more robust assessment of the stability and generalization performance of the GA-based search than reliance on a single optimization run. Across the five independent runs, test accuracy ranged from 85.71% to 89.46%, with a mean of 87.89%. The difference between the best validation fitness and the corresponding independent test accuracy ranged from 3.68 to 6.29 percentage points, with a mean generalization gap of approximately 5.12 percentage points. This moderate and relatively consistent gap indicates that validation fitness provides a somewhat optimistic estimate of performance on unseen data, which may be expected when candidate configurations are selected according to their validation performance. Importantly, the independent test set was not used for chromosome evaluation, fitness computation, or selection during the genetic search and was used only for the final assessment of the selected solution. Consequently, the reported test accuracy constitutes an independent estimate of the generalization performance of the final model. Given the limited search budget ((G = 2)), future experiments with a larger number of generations and systematic investigation of population size, mutation rate, crossover parameters, and population diversity could determine whether further performance improvements are achievable.

Fig. 3

Best validation fitness reached during the genetic search (dark blue) versus the corresponding independent test accuracy of the final retrained chromosome (light blue), for each of the five independent runs (seeds 0–4). A consistent generalization gap of approximately 3.7–6.3 percentage points (mean 5.12 points) is observed between validation fitness and test accuracy across all five runs. This moderate and stable gap indicates that the validation-based fitness function provides a slightly optimistic but reliable estimate of generalization performance, and that the search does not overfit the independent test set, which is used only once per run for the final evaluation.

Computational complexity analysis

The computational complexity of the proposed GA-based optimization procedure mainly depends on the cost of training and evaluating each candidate model. In the proposed framework, each chromosome represents one candidate convolutional configuration. Therefore, the total computational cost is determined by the number of individuals in the population, the number of generations, the number of training epochs used to evaluate each individual, and the training cost of each candidate model.

For a convolutional layer, the approximate computational complexity can be expressed as:

$$\begin{aligned} \mathscr {O}(H \times W \times C_{\text {in}} \times C_{\text {out}} \times K^2) \end{aligned}$$

(1)

where H and W denote the height and width of the feature map, \(C_{\text {in}}\) is the number of input channels, \(C_{\text {out}}\) is the number of output filters, and K is the kernel size. Since the genetic algorithm modifies both the number of filters and kernel sizes in the convolutional blocks, these parameters directly influence the computational cost of each candidate configuration.

The overall computational complexity of the GA-based optimization process can be expressed as:

$$\begin{aligned} \mathscr {O}(N \times G \times E \times T_{\text {epoch}}) \end{aligned}$$

(2)

where N is the population size, G is the number of generations, E is the number of epochs used to train each individual, and \(T_{\text {epoch}}\) represents the average computational cost of one training epoch for a candidate model. In the reported experiment, the parameters were set as: \(N = 5\), \(G = 2\), and \(E = 20\). Therefore, the computational cost increases linearly with the number of individuals, generations, and training epochs.

This analysis shows that the GA-based optimization procedure is computationally more expensive than training a single fixed model because multiple candidate configurations must be trained and evaluated. However, it provides a systematic search mechanism that reduces dependence on manual architectural tuning and allows exploration of different convolutional configurations in a structured manner. Given the limited number of generations used in this study, future work should investigate larger populations, more generations, early stopping strategies, and parallel evaluation of individuals to improve both optimization reliability and computational efficiency.

Results and discussions

After preprocessing the dataset, we constructed an initial ResNet-50 model. We then proceeded with a systematic hyperparameter tuning process, focusing on one hyperparameter at a time. This approach allowed us to explore a range of values for each hyperparameter while evaluating the model performance using a validation set. Once we identified a value that yielded optimal performance for a specific hyperparameter, we fixed it and proceeded to optimize the next hyperparameter. This iterative process continued until we observed no significant improvement in the model accuracy, indicating that we had reached an optimal configuration for the current set of hyperparameters. Finally, we compared the performance of different hyperparameter configurations to identify the values that resulted in the best overall accuracy for the classification of dental caries.

The following elements summarize the two key components of our tuning process:

  • Number of epochs: An epoch refers to the number of times the entire dataset is passed through the model during training. In our experiments, we tested several epoch values ranging from 10 to 50 to determine the optimal number for the network to converge effectively.

  • Sigmoid activation function: The sigmoid activation function is commonly used in neural networks for gradient backpropagation. It introduces non-linearity, which is essential for learning complex patterns. We employed the sigmoid function in our work due to the binary nature of our classification task, which involves two classes (caries and non-caries).

Evaluation metrics

The performance of the proposed model was evaluated using several standard classification metrics. Since dental caries detection is formulated as a binary classification problem, the positive class refers to caries images, whereas the negative class refers to non-caries images. Accordingly, true positives (TP) represent caries images correctly classified as caries, true negatives (TN) represent non-caries images correctly classified as non-caries, false positives (FP) represent non-caries images incorrectly classified as caries, and false negatives (FN) represent caries images incorrectly classified as non-caries.

Accuracy measures the overall proportion of correctly classified samples and is defined as:

$$\begin{aligned} \text {Accuracy} = \frac{TP + TN}{TP + TN + FP + FN} \end{aligned}$$

(3)

Precision measures the proportion of images predicted as caries that are truly caries and is defined as:

$$\begin{aligned} \text {Precision} = \frac{TP}{TP + FP} \end{aligned}$$

(4)

Recall, also referred to as sensitivity, measures the proportion of actual caries cases correctly detected by the model and is defined as:

$$\begin{aligned} \text {Recall} = \frac{TP}{TP + FN} \end{aligned}$$

(5)

Specificity measures the proportion of non-caries cases correctly identified and is defined as:

$$\begin{aligned} \text {Specificity} = \frac{TN}{TN + FP} \end{aligned}$$

(6)

The F1-score represents the harmonic mean of precision and recall and is defined as:

$$\begin{aligned} \text {F1-score} = 2 \times \frac{\text {Precision} \times \text {Recall}}{\text {Precision} + \text {Recall}} \end{aligned}$$

(7)

The area under the receiver operating characteristic curve (AUC-ROC) is also used to evaluate the model’s ability to discriminate between caries and non-caries images across different classification thresholds.

These metrics are appropriate for dental caries detection because they assess complementary aspects of diagnostic performance. High recall is important because it reduces the risk of missed caries cases, which may delay treatment. High precision is also important because it reduces false positive predictions and therefore limits unnecessary follow up examinations or interventions. Specificity further evaluates the ability of the model to correctly identify non-caries cases.

Experimental settings

The experiments were conducted using a publicly available dental caries dataset obtained from the Kaggle repository. The dataset consisted of 1, 554 color dental images belonging to two classes: caries and non-caries. The dataset was divided into training and testing subsets. The training subset contained 1, 260 images, including 945 caries images and 315 non-caries images, whereas the testing subset contained 294 images, including 210 caries images and 84 non-caries images. It should be noted that the dataset used in this study represents a controlled benchmark dataset obtained from a public Kaggle repository. Therefore, it may not fully capture the variability and complexity of real clinical radiographic data. In particular, the dataset does not explicitly cover the full range of real world artifacts and acquisition conditions that may occur in dental imaging, such as metallic restorations, crowns, orthodontic brackets, saliva-related noise, food debris, varying exposure levels, motion blur, and device-dependent image quality differences. Consequently, the reported performance should be interpreted within the limits of this dataset.

Before training, all images were resized to \(224 \times 224\) pixels to ensure compatibility with the input requirements of the deep-learning model. Pixel values were normalized to reduce the effect of illumination variation and to improve training stability. To address the imbalance between caries and non-caries samples (945 and 315 images, respectively), class weights were computed from class frequencies using a balanced weighting scheme:

$$\begin{aligned} w_c = \frac{n_{\textrm{samples}}}{n_{\textrm{classes}} \times n_{\textrm{samples},c}} \end{aligned}$$

(8)

as implemented by the compute_class_weight utility from scikit-learn, and incorporated directly into the loss function during training. No oversampling or undersampling was performed; instead, the loss contribution of the minority (non-caries) class was increased to reduce bias toward the majority class. These class weights were applied consistently across all GA individuals during training.

To reduce the risk of data leakage, the dataset was divided into training and testing subsets before model training and evaluation. Data augmentation was applied only to the training subset, while the test subset was kept unchanged and used exclusively for final evaluation. In addition, duplicate and near-duplicate images were checked to minimize the possibility that visually identical or highly similar samples appeared in both training and testing subsets. These precautions were adopted to ensure that the reported performance reflects model generalization on the held-out test set rather than leakage from the training process.

To improve generalization and reduce overfitting, data augmentation was applied to the training images. The augmentation operations included horizontal flipping, random rotation within \(\pm 15^\circ\), and random zoom within \(\pm 10\%\). No augmentation was applied to the test set, which was kept separate and used only for final model evaluation.

The proposed model was trained using the Adam optimizer. Each candidate configuration generated by the genetic algorithm was trained for 20 epochs. The classification accuracy was used as the fitness criterion for ranking candidate configurations during the GA optimization process. The final optimized model was evaluated using accuracy, precision, recall, F1-score, confusion matrix, and area under the ROC curve.

The experiments were implemented using Python 3.9, TensorFlow 2.10, and Keras 2.10. The computations were performed on a workstation equipped with an Intel Core i7 CPU, 16 GB RAM, and an NVIDIA GPU with CUDA support. Five independent runs were performed using random seeds \(0-4\). Seed 0 was prespecified as the primary run and is used for the confusion matrix reported in Fig. 7. The results across the five runs are reported as mean ± standard deviation in Table 9. The same experimental protocol, including data splitting, population initialization, model training, and evaluation, was applied independently to each run.

To assess the stability of the reported results across different experimental conditions, the full experiment was repeated five times using different random seeds rather than performing k-fold cross-validation, which would require retraining multiple GA optimization processes and is computationally prohibitive given the iterative nature of the proposed framework. The proposed ResNet-GA model achieved a mean test accuracy of 87.89% ± 1.41% across the five independent runs. The corresponding mean macro-precision, macro-recall, and macro-F1-score were 89.74% ± 0.95%, 80.38% ± 3.04%, and 83.32% ± 2.58%, respectively.

Implementation of the GA-based optimization process

This subsection describes the experimental setup used to evaluate the proposed ResNet-GA model. We began by preprocessing the dental image dataset and training the baseline ResNet-50 model. To enhance its performance, we applied a genetic algorithm (GA) to optimize several key hyperparameters.

Each individual in the GA population encodes a ResNet-50 configuration. The chromosome is composed of three genes, each of which represents a convolutional layer. Each gene stores two values: the number of filters and kernel size. For instance, chromosome 68, 15; 68, 18; 10, 6 corresponds to a network with three layers, where the first two layers have 68 filters with \(15 \times 15\) and \(18 \times 18\) kernels, and the third layer has 10 filters with a \(6 \times 6\) kernel.

We initialized the GA by using a population of five randomly generated individuals. Each ResNet50 configuration was trained for 20 epochs, and its validation accuracy was used as the fitness value. Selection was performed using a fitness based method and crossover combined parent chromosomes by taking the first half of the genes from parent 1 and the second half from parent 2. A fixed mutation rate is then applied to introduce variability into the offspring.

The detailed configurations of the randomly generated initial population are summarized in Table 5. The evolution of the population across two generations allowed us to track the performance improvement and identify the best performing model

Table 5 Illustrative randomly generated initial population, used to describe the chromosome encoding and the crossover and mutation operators. The validation fitness values obtained during the actual evolutionary search are reported in Table 9, where the best validation fitness across the five independent runs ranged from 90.87% to 94.44%.

Full size table

The initial population consisted of five ResNet-50 models (see Table 5), each considered an individual whose validation accuracy serves as the fitness value. This fitness value is critical for evaluating the performance of each model in the context of dental caries classification. To select individuals for mutation, we choose the top n individuals with the highest fitness values, where n is the desired number for subsequent mutation (see Table 6).

Table 6 Five ResNet-50 parameters.

Full size table

These selected individuals were referred to as ResNet-50 n°1, ResNet-50 n°2, etc. We employ a fixed crossover strategy, where the first half of the genes are inherited from parent 1 and the remaining genes are taken from parent 2. This method ensures a consistent and controlled exchange of genetic information between parents, which is essential for maintaining structural integrity while promoting diversity within the offspring population. The resulting off- spring populations are presented in Table 7.

This structured crossover approach allows for exploring new architectures while leveraging the strengths of the parent models.

Table 7 Population of children.

Full size table

In our crossover implementation, parent 1 exchanges the number of filters in layer 3 with the number of filters from parent 2, and the same operation is performed for the filter size in layer 3. This targeted exchange of hyperparameters was designed to enhance the offspring’s ability to learn from the strengths of both parents, potentially leading to improved accuracy in the classification task.

ResNet-50 Parent 1: 68 68 10, 15 18 6

ResNet-50 Parent 2: 45 48 65, 2 7 8

ResNet-50 Child 1: 68 68 65, 15 18 8

ResNet-50 Child 2: 45 48 10, 2 7 6

ResNet-50 Child 3: 68 68 65, 15 18 8

Table 8 shows the population of mutated children.

Table 8 Population of mutated children.

Full size table

ResNet-50 Child 1: 68 68 65, 15 18 8 becomes:

ResNet-50 Child 1: 72 68 70, 18 20 8 And so on for the other two children.

After generating a new offspring population, the steps of selection, crossover, and mutation are repeated to create subsequent generations. This iterative process continued until Generation 2. During this phase, we closely monitored the performance metrics of the models, particularly their accuracy in the dental caries classification task. We observed that the improvement in the results plateaued beyond this point, indicating that further iterations yielded diminishing returns regarding model performance.

Consequently, we decided to use Generation 2 as the stopping criterion for our genetic algorithm.

The decision to terminate the evolutionary process after Generation 2 was based on the observed stabilization of the best fitness value within the adopted computational budget, rather than on a formal demonstration of genetic algorithm convergence. Since each fitness evaluation requires training and validating a deep neural network, extending the evolutionary search substantially increases computational cost. Generation 2 was therefore selected as a practical stopping point for evaluating the feasibility of the proposed optimization framework under the adopted experimental settings.

Table 9 Individual-run and aggregated performance of the proposed ResNet-GA framework over five independent runs (seeds 0–4). \(^{\dagger }\)Seed 0 is the prespecified primary run reported in Fig. 7.

Full size table

The performance reported in Table 9 represents the mean ± standard deviation over five independent runs and therefore reflects the variability of the proposed model across different random initializations. The independent test set consisted of 294 images and was used only once for the final evaluation of the winning chromosome in each run. The primary run (seed 0), prespecified before the five-run evaluation, achieved a test accuracy of 85.71%, with the confusion matrix reported in Fig. 7. The corresponding results for all five independent runs are reported in Table 9.

Convergence analysis

To evaluate the stability of the proposed GA-based optimization procedure, we analyzed the best validation fitness reached by each of the five independent runs, alongside the corresponding independent test accuracy of the final retrained chromosome (Fig. 3). Across all five runs, the best validation fitness ranged from 90.9% to 94.4% (mean 93.0%), while the corresponding test accuracy ranged from 85.7% to 89.5% (mean 87.9%, Table 9). A consistent generalization gap of 3.7–6.3 percentage points (mean 5.12 points) was observed between validation fitness and test accuracy in every run, indicating that the validation-based fitness function generalizes reasonably well to unseen data, and that the magnitude of the optimism introduced by model selection on the validation subset remains limited and stable across runs.

Because the genetic search was limited to two generations per run \((G = 2)\) for computational reasons, a fine-grained generation by generation convergence trace is not reported here. Instead, aggregating results across five independent runs provides a more statistically grounded assessment of convergence behavior than a single trace would, at the cost of not characterizing within-run, generation-by-generation dynamics. A more detailed convergence analysis (tracking validation fitness at each generation across multiple independent runs) is identified as a direction for future work.

Statistical validation

To examine whether the performance difference between the proposed ResNet-GA model and the baseline ResNet-50 model was statistically supported, we compared their paired predictions on the held-out test set using McNemar’s test (Table 10), which evaluates classification disagreement at the instance level. For the primary run (seed 0), the contingency table of paired predictions was: 250 samples correctly classified by both models, 0 correctly classified only by the baseline, 2 correctly classified only by ResNet-GA, and 42 misclassified by both (total \(n = 294\); baseline accuracy \(= 85.03\%\), ResNet-GA accuracy \(= 85.71\%\) on the primary run). McNemar’s exact test on this table yielded \(p = 0.5000\), indicating no statistically significant difference between the two models’ classification behavior at \(\alpha = 0.05\). This outcome is consistent with the small size of the independent test set and with the moderate effect size observed on the primary run (+0.68 percentage points), and is reported here for transparency rather than omitted. It should be noted that this test is restricted to the prespecified primary run (seed 0), for which the accuracy gap over the baseline was smallest (+0.68 percentage points) among the five independent runs (Table 9, range +0.68 to +4.43 percentage points). Per-sample paired predictions were not persisted for the baseline model across the remaining four seeds, so McNemar’s test could not be computed for those runs; consequently, we cannot rule out that the larger accuracy gaps observed on seeds 1–4 would correspond to a statistically significant instance-level difference. This test should therefore be interpreted as evidence specific to the primary run rather than as a general claim of statistical equivalence across all runs. The consistent advantage of ResNet-GA is better evidenced by the multi-seed evaluation, where the mean accuracy of 87.89% ± 1.41% exceeds the baseline’s 85.03% across all five independent runs.

We had also intended to report a Wilcoxon signed-rank test comparing paired per-run performance between the two models across independent seeds, as this non-parametric test does not require the assumption of normality and is appropriate for comparing two related models evaluated on matched samples. However, a valid paired comparison requires matched samples obtained under the same experimental conditions for both models across multiple independent runs. In the present experimental setup, only the ResNet-GA model was evaluated across five independent seeds (Table 9); the baseline ResNet-50 model was trained and evaluated only once, since its transfer learning head involves no stochastic architecture search.

Table 10 Statistical validation of the proposed ResNet-GA model (corrected). \(^*\)The baseline model was evaluated only once (deterministic transfer-learning head); a valid paired Wilcoxon test requires repeated baseline runs under matched random seeds, which are not yet available. \(^\dagger\)McNemar’s test was computed only for the prespecified primary run (seed 0); it was not computable for seeds 1–4 because per-sample baseline predictions were not persisted for those runs.

Full size table

Ablation study

To evaluate the individual contribution of each core component of the proposed framework, an ablation study was conducted by progressively introducing the main architectural and optimization elements. Four distinct configurations were compared under identical evaluation conditions: (1) baseline ResNet-50 with randomly initialized weights (no transfer learning, no search), (2) ResNet-50 with ImageNet transfer learning, (3) ResNet-50 with transfer learning combined with random search for filter and kernel optimization under an equivalent search budget, and (4) the proposed ResNet-GA framework.

The ablation results are summarized in Table 11. The baseline ResNet-50 trained from scratch achieved an accuracy of 71.43% (AUC = 0.762), reflecting the difficulty of training a deep architecture on a limited medical imaging dataset without prior feature representations. Introducing ImageNet-based transfer learning provided a substantial performance gain of +13.60 percentage points in accuracy (85.03%, AUC = 0.895), confirming that pretrained visual representations are essential for dental caries detection.

Hyperparameter optimization via random search further improved accuracy to 86.73% (+1.70 percentage points), demonstrating the utility of searching for task-specific layer configurations. Finally, the proposed ResNet-GA model achieved the highest performance across all evaluation metrics, reaching an accuracy of 87.89% (precision = 89.74%, recall = 80.38%, F1-score = 83.32%, AUC = 0.926). This represents a +1.16 percentage point improvement over random search and a +2.86 percentage point improvement over the standard transfer-learning baseline. These results confirm that while transfer learning provides the foundational feature extraction capability, genetic algorithm-driven evolutionary optimization effectively refines block-level hyperparameters to maximize classification performance.

Table 11 Ablation study of the proposed ResNet-GA framework showing the progressive contribution of transfer learning and evolutionary optimization on dental caries classification.

Full size table

These results demonstrate that the proposed ResNet-GA framework benefits from both transfer learning and evolutionary optimization. Although transfer learning provides a substantial improvement over the baseline model, the best performance is obtained when GA-based optimization is applied. This confirms the added value of the genetic algorithm in identifying more effective model configurations for dental caries classification.

Comparison with state of the art methods

To provide a rigorous and fair evaluation of the proposed ResNet-GA framework, we conducted a controlled within-dataset benchmark comparing it against representative deep learning architectures and hybrid classifiers. All competing models were evaluated on the exact same dental caries dataset, using identical training/testing splits (with stratified partitioning), standard \(224 \times 224\) image preprocessing, and a consistent decision threshold of 0.5. This strict protocol ensures that observed performance variations reflect architectural and optimization efficacy rather than discrepancies in data distribution or experimental design.

Table 12 summarizes the performance of the proposed ResNet-GA framework in comparison with baseline transfer-learning ResNet-50, DenseNet-121, and a hybrid ResNet-50 combined with an XGBoost classifier. As shown in the table, the proposed ResNet-GA model achieved superior results across all evaluated performance metrics, obtaining an accuracy of 87.89%, a macro-averaged precision of 89.74%, a macro-averaged recall of 80.38%, an F1-score of 83.32%, and an AUC of 0.926. In comparison, the DenseNet-121 model attained an accuracy of 86.05% (AUC = 0.904), while the ResNet-50+XGBoost hybrid reached 86.39% (AUC = 0.909). The standard transfer-learning ResNet-50 baseline yielded 85.03% accuracy (AUC = 0.895).

These empirical findings demonstrate that evolutionary search over block-level filter counts and kernel sizes effectively tailors the feature extraction hierarchy to the subtle morphological patterns of dental caries, providing a measurable advantage (+1.50% to +2.86% in accuracy and +0.017 to +0.031 in AUC) over standard off-the-shelf convolutional backbones and gradient-boosted tabular heads.

In addition to this controlled within-dataset benchmark, external findings from previously published literature are compiled in Table 13 for contextual purposes. Consequently, the controlled within-dataset results reported in Table 12 provide the most reliable evidence of the effectiveness of the proposed framework.

Table 12 Comparative performance of the proposed ResNet-GA framework against baseline and competing architectures on the dental caries dataset.

Full size table

Although the proposed ResNet-GA model outperformed the compared baselines under the same experimental conditions, cross-study comparisons with previously published works should be interpreted cautiously. Differences in dataset size, image quality, annotation protocols, preprocessing methods, and evaluation strategies can substantially affect reported performance. Therefore, the within dataset comparison provides the most reliable evidence of the effectiveness of the proposed framework.

Interpretation of the results

Firstly, we trained the baseline transfer-learning ResNet-50 model on the dental caries imaging dataset. After analyzing the experimental results obtained from the training trajectories, multi-run evaluations, confusion matrices, and ROC analyses, we deduced the following observations:

Training and Validation Dynamics: Figure 4 illustrates the relationship between accuracy, loss, and the number of training epochs for both the training and validation/testing sets. As the number of epochs increased, we observed a general trend of increasing accuracy (stabilizing between 74% and 77%) accompanied by a steady decrease in loss (declining from \(\approx\) 0.57 to below 0.49). This trajectory confirms that the model successfully learned informative visual patterns from the training data. The oscillatory behavior observed in the validation accuracy and loss curves of the baseline ResNet-50 model can be attributed to two main factors: first, the default learning rate of 0.001 during fine-tuning on a relatively small dataset (1,260 images) can introduce gradient update sensitivity; second, with a batch size of 32 (approximately 39 mini-batch updates per epoch), stochastic variations in batch composition induce fluctuations. These empirical instabilities directly motivated the GA-based optimization in the proposed ResNet-GA framework, which systematically searched for more effective block-level convolutional configurations. It should be noted that the accuracy values shown in Figure 4 reflect the per-epoch training dynamics under data augmentation, which artificially increases the difficulty of the training batches. The final evaluation of the same baseline model on the held-out test set, reported in Table 11 and table 12, yields an accuracy of 85.03%, which is consistent with the upward trend observed at the end of the fine-tuning process.

Figure 4

Training dynamics of the baseline transfer-learning ResNet-50 over 20 epochs. Left: training and testing accuracy. Right: training and testing loss. The curves correspond to the baseline configuration used as the reference model in the ablation study (Table 11).

Multi-Seed Stability and Confusion Matrix Analysis: Figure 5 presents the confusion matrices across five independent evolutionary runs of the ResNet-GA model (Seeds 0 to 4, evaluated on the held-out test set of \(n=294\) images comprising 210 caries and 84 non-caries samples):

  • Seed 0 (Primary Run): 208 True Positives (TP), 44 True Negatives (TN), 40 False Positives (FP), and 2 False Negatives (FN), achieving an accuracy of 85.71%.

  • Seed 1: 205 TP, 53 TN, 31 FP, and 5 FN, achieving an accuracy of 87.76%.

  • Seed 2: 202 TP, 59 TN, 25 FP, and 8 FN, achieving an accuracy of 88.78%.

  • Seed 3: 206 TP, 57 TN, 27 FP, and 4 FN, achieving an accuracy of 89.46%.

  • Seed 4: 207 TP, 51 TN, 33 FP, and 3 FN, achieving an accuracy of 87.76%.

Fig. 5

Confusion matrices on the independent test set (\(n = 294\)) for each of the five independent ResNet-GA runs (seeds 0–4), each having selected its own best chromosome through a genetic search evaluated exclusively on a held-out validation split. These five matrices illustrate the run-to-run variability underlying the mean ± standard deviation reported in Table 9.

Across all five independent runs, the ResNet-GA model demonstrated remarkably high and consistent sensitivity on the caries class (TP ranging from 202 to 208 out of 210, with only 2 to 8 false negatives across all seeds). This low false-negative rate is of paramount clinical importance in dental diagnostics, where missing an active carious lesion carries substantial clinical risk. Furthermore, the overall test accuracy remained highly consistent (87.89% ± 1.41%, ranging from 85.71% to 89.46%), demonstrating that evolutionary optimization converges to robust and reproducible architectural solutions despite random population initializations.

ROC Curve and Discriminative Capacity: Figure 6 illustrates the Receiver Operating Characteristic (ROC) curves comparing the proposed ResNet-GA model against the baseline ResNet-50 transfer-learning model. The Area Under the Curve (AUC) achieved by ResNet-GA was 0.926, compared with 0.895 for the baseline ResNet-50. The ROC curve of ResNet-GA consistently dominates that of the baseline model across the entire False Positive Rate range, demonstrating superior diagnostic discrimination. The steep initial ascent of the ResNet-GA curve indicates that the model captures a large proportion of true positive cases at very low false positive thresholds, confirming its suitability as an automated screening tool.

Fig. 6

ROC curve comparison between the baseline ResNet-50 and the proposed ResNet-GA framework on the test dataset.

Primary Run Confusion Matrix: Figure 7 provides a comprehensive breakdown of the primary ResNet-GA model (Seed 0) evaluated on the held-out test dataset (\(n=294\)). The confusion matrix details 208 True Positives (TP), 44 True Negatives (TN), 40 False Positives (FP), and only 2 False Negatives (FN), resulting in an overall classification accuracy of 85.71%. At the class-specific level, the model achieved an exceptional sensitivity (recall) of 99.05% on the positive caries class (\(\frac{208}{210}\)) alongside a Negative Predictive Value (NPV) of 95.65% (\(\frac{44}{46}\)). This means that only 2 out of 210 true carious lesions were missed across the entire test cohort. In clinical dental screening, minimizing false negatives is of paramount diagnostic importance, as undetected caries can progress rapidly to irreversible pulp damage, periapical pathology, and severe tooth loss. While the model exhibited 40 false positives (yielding a specificity of 52.38% and a precision of 83.87% for caries detection), such benign false alarms can be readily identified and ruled out during routine confirmatory dental examination with minimal clinical consequence. Overall, the primary run achieved a macro-averaged precision of 89.76%, a macro-averaged recall of 75.71%, and a macro F1-score of 79.26%, underscoring the model’s reliability as an automated first-line diagnostic screening assistant.

Fig. 7

Confusion matrix of the prespecified primary run (seed 0) on the independent test set(n = 294). The matrix [[208, 2], [40, 44]] contains 252 correct predictions, corresponding to a test accuracy of 85.71%. This run was designated as the primary run for Fig. 7 before the five independent runs were executed, to avoid post-hoc selection.

Comparison of results

Table 13 presents a comparative overview of the proposed ResNet-GA framework and representative deep learning approaches reported in the literature for dental caries detection and classification. The comparison includes conventional convolutional neural networks (CNNs), transfer learning approaches, object detection frameworks, hybrid architectures, and optimization-driven methods. However, direct comparisons between studies should be interpreted with caution because existing approaches were evaluated under different experimental conditions, including variations in dataset size, image acquisition modality, preprocessing procedures, annotation strategies, and evaluation protocols. Therefore, the reported results are considered contextual benchmarks rather than a strict ranking of model performance.

The proposed ResNet-GA framework achieved an average accuracy of 87.89% ± 1.41%, a macro-averaged precision of 89.74% ± 0.95%, a macro-averaged recall of 80.38% ± 3.04%, and an F1-score of 83.32% ± 2.58% across five independent evolutionary runs (with an AUC of 0.926 on the primary evaluation run). These results demonstrate the effectiveness of evolutionary optimization for tailoring the convolutional representation of a ResNet-50 backbone to medical dental imaging. Unlike conventional approaches relying on manually selected hyperparameters or standard off-the-shelf transfer learning, the proposed framework introduces a genetic algorithm-based optimization strategy that automatically searches for task-optimal filter counts and kernel sizes while preserving the structural advantages of the residual architecture.

Early transfer learning-based approaches demonstrated the effectiveness of pretrained CNN architectures for dental caries detection.The authors of18 investigated ResNet50 and MobileNet models using a dataset containing 395 normal and 395 caries images and reported an accuracy of 96.12%. Although the obtained performance was promising, the relatively limited dataset size may affect the generalization capability of the model. Similarly, the authors of19evaluated several pretrained CNN architectures, including VGG-16, VGG-19, DenseNet-121, and Inception V3, and reported a maximum accuracy of 98.3% using VGG-16. These findings confirm the relevance of transfer learning for dental image analysis. Although these previous studies reported high accuracies on smaller, curated datasets (e.g., 96.12% in18 and 98.3% in19), direct numerical comparisons across different datasets are inherently limited. Unlike those works which relied on fixed architectures and manual tuning, ResNet-GA introduces an automated evolutionary search that optimizes convolutional representations directly on a standardized benchmark while reducing empirical trial-and-error.

The modified DenseNet-121 approach proposed by the work of24 achieved an accuracy of 95% for oral pathology classification. While the results confirm the suitability of DenseNet-based feature extraction, the limited dataset size may affect the generalization of the model.

The CNN-based framework proposed by the work of22 reported very high classification performance for dental caries and other oral conditions. Nevertheless, the evaluation was conducted on a specific dataset, and additional validation on more diverse clinical data would be required to assess generalization capability.

Hybrid learning strategies have also been explored for dental caries detection. The research in21 combined ResNet-50 feature extraction with XGBoost classification and achieved an accuracy of 97.81%. Although hybrid approaches can improve predictive performance, their combination of multiple learning components may increase computational complexity. The proposed ResNet-GA framework follows a different strategy by optimizing the internal configuration of a single CNN architecture, providing a more controlled and reproducible optimization process.

Object detection approaches have recently gained attention in dental applications. The study of20 proposed a YOLO-based framework combined with transfer learning architectures, including ResNet50, ResNet101, VGG16, AlexNet, and DenseNet, achieving an accuracy of 86.96% using non-standardized smartphone photographs. The lower performance compared with radiographic approaches highlights the influence of image acquisition conditions and variability. More recent YOLO-based methods further extended this direction. The work of31 evaluated YOLOv8x for clinical caries detection using intraoral photographs, while the work of33 investigated YOLOv9 for dental X-ray analysis. These studies demonstrate the potential of real-time object detection systems; however, they mainly focus on lesion localization and detection rather than optimizing CNN feature extraction for classification. The Faster R-CNN framework combined with ResNet-50 and ResNet-101 proposed by the authors of23 demonstrated the effectiveness of object detection approaches for dental caries localization. However, object detection and classification tasks involve different optimization objectives, making direct comparison with classification-based approaches such as ResNet-GA difficult. The ViT and U-Net hybrid framework proposed by the study of25 achieved an accuracy of 99%, demonstrating the potential of combining transformer-based feature extraction with segmentation architectures. However, the increased architectural complexity may introduce additional computational requirements compared with more focused optimization strategies.

Recent advanced architectures have also been investigated for dental image analysis. The researchers of28 proposed a transformer-based detection model for early dental caries detection using panoramic radiographs. Transformer architectures provide powerful global feature modeling capabilities and have shown promising sensitivity for subtle lesions. Nevertheless, they generally require larger annotated datasets and higher computational resources compared with conventional CNN architectures. Furthermore, the authors of29 demonstrated that CNN-based architectures remained highly competitive compared with transformer and hybrid architectures for dental image segmentation tasks, suggesting that increased model complexity does not always guarantee improved performance.

Generative approaches have recently been proposed to overcome limitations related to insufficient dental image datasets. The work of30 introduced Pulp-Caries-GAN, a GAN-based framework combined with biomimetic metaheuristic optimization for pediatric caries segmentation, achieving a Dice coefficient of 95.12%. Although this approach effectively improves segmentation through synthetic data generation, the integration of additional generative components increases the complexity of the overall system. In contrast, ResNet-GA improves classification performance through evolutionary optimization of convolutional parameters without requiring synthetic image generation.

Recent studies have also explored alternative diagnostic modalities beyond conventional radiographic imaging. The study of35 proposed a PS-OCT-based method for early dental caries detection and quantitative analysis, while the authors of34 investigated machine learning algorithms based on spectrophotometric CIELAB measurements for occlusal caries severity classification. These approaches provide valuable complementary diagnostic information; however, they rely on specialized acquisition systems that differ from the conventional dental image classification scenario addressed in the present study.

Large-scale benchmarking studies and systematic reviews have further emphasized the challenges associated with comparing dental AI models. The research by27 introduced the MD-OPG benchmark for pediatric dental caries detection using panoramic radiographs, highlighting the importance of standardized datasets and evaluation protocols. The authors of37 performed a modality-stratified systematic review and meta-analysis of Faster R-CNN and Mask R-CNN approaches, demonstrating that reported performance varies significantly according to dataset characteristics, annotation procedures, and imaging modalities. Similarly, the work of38 emphasized the gap between algorithmic performance and clinical translation, indicating that high benchmark accuracy alone is insufficient to guarantee clinical applicability.

Considering the diversity of existing approaches, the contribution of ResNet-GA should not be interpreted solely through direct numerical comparison with previous models. Instead, its main contribution lies in introducing a focused evolutionary optimization strategy that improves a well-established CNN backbone while preserving architectural stability and interpretability. Unlike neural architecture search approaches that attempt to generate completely new network structures, ResNet-GA performs targeted optimization of selected convolutional parameters, providing a balance between performance improvement and computational feasibility.

Overall, the obtained results indicate that genetic algorithm-based optimization represents a promising strategy for enhancing deep learning models in dental caries detection. Nevertheless, further validation using larger multi-center datasets, standardized benchmarks, and external clinical evaluations remains necessary to assess the generalization capability and practical applicability of the proposed framework.

However, these results should not be interpreted as definitive evidence of universal superiority over all existing dental AI systems, because recent approaches based on vision transformers, YOLO detectors, GAN-based augmentation, and alternative radiographic modalities address distinct diagnostic objectives and are evaluated on heterogeneous datasets under varied protocols. Therefore, the main contribution of ResNet-GA lies in introducing a lightweight and computationally controlled evolutionary optimization strategy that enhances a well-established deep learning architecture while maintaining structural stability and interpretability.

In this context, the proposed framework should be interpreted as a targeted, computationally efficient evolutionary hyperparameter optimization strategy rather than a full neural architecture search (NAS) system. Its primary contribution is to demonstrate that targeted genetic optimization of block-level convolutional hyperparameters can systematically enhance classification performance while fully preserving the structural integrity and convergence stability of a proven residual backbone.

Table 13 Comparison of the proposed ResNet-GA framework with representative deep learning approaches for dental caries detection and analysis. Reported performances are provided for contextual comparison because differences in datasets, imaging modalities, and evaluation protocols limit direct ranking between studies.

Full size table

Future research directions

Future research should address the methodological and clinical limitations identified in this study through realistic and technically justified extensions.

First, the proposed GA-based optimization framework could be extended toward multi-objective optimization. Instead of optimizing classification performance alone, future models could jointly consider accuracy, computational cost, sensitivity, specificity, and false positive rate. This direction is clinically relevant because reducing false positives may help avoid unnecessary follow up procedures, while maintaining high sensitivity remains important for minimizing missed caries cases.

Second, the proposed framework should be evaluated on additional dental imaging modalities, such as cone-beam computed tomography (CBCT) and intraoral scan datasets. Such validation would help assess the robustness of the model across different acquisition protocols, image resolutions, anatomical structures, and clinical imaging conditions.

Third, explainability methods such as Grad-CAM should be integrated to improve the interpretability of model predictions. Visual explanations may help clinicians verify whether the model focuses on clinically meaningful dental regions rather than irrelevant artifacts or background patterns.

Finally, federated learning could be investigated as a privacy preserving strategy for training models across multiple clinics or institutions without directly sharing patient data. This direction is relevant because access to large and diverse clinical datasets is often limited by privacy regulations and data governance constraints in healthcare.

The high accuracy obtained in this study should also be interpreted cautiously because the dataset contains only 1,554 images and may represent a relatively controlled benchmark scenario. Such datasets can sometimes contain clear visual differences between classes, limited acquisition variability, or repeated visual patterns, which may lead to higher performance than would be expected in real clinical settings. Although precautions were taken to reduce data leakage and overfitting, the limited dataset size means that the risk of overestimating generalization performance cannot be fully excluded. Therefore, the reported results should be considered promising within the evaluated dataset rather than definitive evidence of clinical robustness. Given the modest size of the test set \((n = 294)\), the reported accuracy should be interpreted with its associated statistical uncertainty rather than as an exact figure; evaluation on larger, independent, multi-center datasets is required before the point estimate can be treated as a stable clinical benchmark.

Conclusion

Artificial intelligence has become an important tool in healthcare, particularly for supporting medical image analysis and reducing the subjectivity associated with manual interpretation. In this study, we addressed the detection and classification of dental caries by developing a hybrid deep learning framework, ResNet-GA, which combines convolutional feature learning with genetic algorithm based optimization. Manual interpretation of dental radiographs can be time consuming and may be affected by inter observer variability, especially when distinguishing early stage caries from benign anomalies, image noise, or artifacts (6566;).

The proposed ResNet-GA framework was designed to improve automated dental caries classification by using ResNet-50 for hierarchical feature extraction and GA-based optimization to search for effective model configurations. The experimental findings showed that the optimized model achieved promising classification performance on the evaluated dataset, with improved accuracy, precision, recall, F1-score, and AUC compared with the baseline configurations considered in this study. These results suggest that evolutionary optimization can contribute to improving the performance of deep learning models for dental image classification.

However, the findings should be interpreted within the limits of the dataset and experimental protocol used in this study. The dataset was relatively limited in size and obtained from a single public source, which may restrict the generalizability of the results to broader clinical populations and different radiographic acquisition conditions. Therefore, although the proposed ResNet-GA framework shows potential as a computer aided tool for dental caries classification, further validation on larger, independent, and multi-center clinical datasets is required before any practical clinical deployment can be considered.

Overall, the proposed ResNet-GA framework provides a promising and experimentally supported contribution to AI assisted dental image analysis. However, its findings remain limited to the evaluated dataset and experimental conditions. Further validation on larger and more diverse clinical datasets is required before considering practical clinical deployment.

来源:Nature 健康护理主题 · nature.com