机器学习预测牙冠颜色变化研究
Evaluation of colour parameters at different preparation depths and prediction of tooth colour using machine learning
该研究使用机器学习技术预测不同预备深度下的牙齿颜色变化,通过测量预备前的牙齿颜色,利用线性回归、多项式回归等模型预测预备后的颜色,以提高美学修复的可预测性。研究共分析了106颗牙齿,结果显示AI模型在预测预备后牙冠颜色方面表现良好。
Introduction
In contemporary dentistry, where aesthetic expectations are continually increasing, treatments are expected to be successful not only functionally but also aesthetically [1]. Preparation depth does not simply refer to the amount of tissue loss in aesthetic restorations of the anterior region. The colour of the abutment is one of the most important factors determining the colour of the final restoration in these treatments [2]. Predicting the shade that may appear on the tooth surface based on the extent of reduction before preparation may be important for planning how to achieve the desired final shade, besides selecting the properties of the ceramic to be used (e.g., thickness and opacity). The use of artificial intelligence (AI) technologies, which are capable of making predictions by learning from datasets regarding tooth colour estimation based on preparation depth, can assist clinicians in pretreatment planning [3,4,5]. Predicting potential colour changes in the abutment tooth before preparation can help prevent unnecessary tissue loss and facilitate achievement of the desired colour outcome. Consequently, restoration thickness can be more accurately planned, thereby improving aesthetic predictability in the final restoration.
In dentistry, the colour space and associated colour coordinates (CIE-L*, CIE-a*, and CIE-b*) defined by the International Commission on Illumination (CIE) are recorded using measurement devices such as spectrophotometers or colourimeters to enable objective colour measurements and ensure a more accurate colour matching process. These values facilitate appropriate mathematical calculations [6].
AI is a field of science aimed at developing systems capable of mimicking human mental activities such as thinking, learning, and decision-making. It has emerged as an extension of efforts to understand how human intelligence works [7, 8]. Human learning is based directly on experiences, whereas the learning process in machines occurs through data. In this context, machine learning is a subfield of AI that allows computers to think and learn on their own [9, 10]. Linear regression (LR), polynomial regression (PR), decision tree regression (DTR), random forest regression (RFR), and support vector regression (SVR) were selected in this study to predict tooth colour based on preparation depth. The rationale for this selection was to test both linear and nonlinear possibilities for the depth–colour relationship and determine which type of mathematical model best fit this relationship [11,12,13,14]. The metrics mean absolute error (MAE), mean squared error (MSE), and coefficient of determination (R²) values were used to evaluate model performance in machine learning [15]. This study aimed to analyze the colour changes occurring at different preparation depths and use these data to predict the likely post-preparation tooth colour by measuring the pre-preparation tooth colour using AI algorithms, thereby supporting evidence-based clinical planning processes.
The aim of this study is to increase predictability regarding the final shade in a subjective field such as aesthetics [16], and, in situations where the extent of preparation can be estimated using methods such as clinical mock-ups [17], to determine the abutment shade prior to preparation [18], thereby enabling the material and its properties to be selected in advance [19]; consequently, to prevent the need for additional appointments or the production of replacement veneers [20], and to achieve a predictable and successful treatment outcome. The first null hypothesis (H0) of this study is “colour values do not change with depth,” whereas the second H0 is “AI models are successful in predicting the abutment colour after preparation, before the preparation is performed.”
Materials and methods
Before starting the study, a power analysis was conducted for a two-way analysis of variance based on a 95% confidence level (1 − α = 0.95) and 95% power (1 – β = 0.95). The analysis showed that including at least 60 samples in the study was sufficient. A total of 106 teeth were included in the study to enhance the reliability of the study.
It was approved by the Ethics Committee for Ethics Committee for Non-Interventional Scientific Research at the Faculty of Medicine, Trakya University (date: January 6, 2025; approval number 01/27). Upper and lower anterior teeth that had been extracted for various reasons and were free of caries, fillings, and root canal treatment were used in the study. The sample included 106 teeth (24 maxillary central incisors, 21 maxillary lateral incisors, 21 maxillary canines, 22 mandibular lateral incisors, and 18 mandibular canines).
An initial colour measurement was then performed. Subsequently, 0.3, 0.5, 0.7, 1.0, and 1.5 mm depth marking burs (Horico FGAS, Berlin, Germany) were used to perform reduction at the specified depths relative to the baseline level under water cooling with an air turbine. The colour measurements were taken at each stage. Color measurements were taken on the same white background; the device was calibrated before each measurement, and each measurement was taken twice at the same depth. [21] A Vita Easyshade V spectrophotometer (Vita Zahnfabrik, Bad Säckingen, Germany) was used for colour measurements. During data collection, the depth of the enamel-dentin junction was recorded for each sample.
Calculation of colour variables
The CIEDE2000 colour difference formula was used for the data obtained from colour measurements. CIEDE2000 formula (CIEDE2000(2:1:1):
$${\Delta E}_{2000}=\sqrt {{\left(\frac{\Delta L^{\prime} }{{k}_{L}{S}_{L}}\right)}^{2}+{\left(\frac{\Delta C^{\prime} }{{k}_{C}{S}_{C}}\right)}^{2}+{\left(\frac{\Delta H^{\prime} }{{k}_{H}{S}_{H}}\right)}^{2}+{R}_{T}\left(\frac{\Delta C^{\prime} }{{k}_{C}{S}_{C}}\right)+\left(\frac{\Delta H^{\prime} }{{k}_{H}{S}_{H}}\right)}$$
This formula used the L*, a*, and b* values from the CIE-Lab* system. In this system, the first measured values are designated as L1*, a1*, and b1*. Measured values obtained from measurements taken at the depths for measurement were recorded and values are designated as L2*, a2*, and b2*. The colour difference between these two measurements was calculated using the CIEDE2000 colour difference formula [22]. In these measurements, the values obtained from the colour measurement performed on the unprepared tooth surface were used as the first measurement values (L1*, a1*, and b1*). These data, recorded at various preparation depths, were analyzed under the following categories: overall colour changes; ΔL, Δa, Δb, and ΔE00 values by preparation depth; ΔL, Δa, Δb, and ΔE00 values relative to the enamel-dentin junction; and evaluation of ΔE00 values according to preparation depth.
AI and machine learning analysis
The colour measurement data were saved in Excel format. The machine learning analyses were performed in the Google Colaboratory (Google Colab) environment. The data were imported using the pandas library. Each row in the dataset contained the corresponding preparation depth along with the associated L*, a*, and b* colour values. The dataset was structured as a long-format table, with each tooth measurement representing a single observation. The columns contained the following variables in order: tooth_id (integers ranging from 1 to 106), depth_mm (0.3 / 0.5 / 0.7 / 1.0 / 1.5 — categorical), L_star (continuous, decimal), a_star (continuous, decimal), b_star (continuous, decimal). A total of 530 rows (106 × 5) were obtained. The dataset was exported in. csv format, retaining the variable names and order, and imported into the Google Colab environment using ‘pd.read_csv‘.”
In the machine learning models, preparation depth was used as the independent variable, whereas L*, a*, and b* values were used as the dependent variables. The dataset was split into an 80% training set and a 20% test set. A sample size of 106 teeth is sufficient, consistent with comparable studies in the machine learning literature [19, 23].
The algorithms used in this study were selected from among the most frequently cited methods in the dental literature for colour and morphology prediction. LR was used to identify simple relationships [8]. LR model was applied, and the model performance was evaluated using the MAE, MSE, and R² metrics. A second-order PR model was applied to examine nonlinear relationships because low R2 values were obtained with the linear model.
“DTR” was selected for hierarchical data clustering, whilst “RDR” and “SVM” were chosen for high-accuracy predictive analyses [24, 25]. Further, DTR, RFR, and SVR models were used to model the relationship between preparation depth and colour components. The performance of all models was compared using MAE, MSE, and R² values.
Statistical analyses
The results of the study were analyzed using IBM SPSS Statistics 27. The normality of the data was assessed using the Kolmogorov–Smirnov and Shapiro–Wilks tests. The data were found to be non-normally distributed. Descriptive statistical methods such as minimum, maximum, mean, standard deviation, median, and frequency were used for evaluating the study data. The quantitative data across the groups were compared using the Kruskal-Wallis test, and Dunn’s test was used for post hoc identification of group differences. The Mann–Whitney U test was used for comparisons between two groups. The chi-square test and the continuity (Yates) correction were used for comparing qualitative data. Significance was assessed at the p < 0.05 level.
Results
The correlation between colour and preparation depth
ΔL, Δa, Δb, and ΔE00 values according to preparation depths:
Statistically significant differences were observed in the ΔL and ΔE00 values across the depths (p = 0.001). Post hoc comparisons revealed that the ΔL and ΔE00 values at 0.3 mm were significantly lower than those at 0.7 mm (p = 0.005), 1.0 mm (p = 0.001), and 1.5 mm (p = 0.001) depths. The ΔL and ΔE00 values at 0.5 mm were significantly lower than those at 1.0 mm (p = 0.006) and 1.5 mm (p = 0.001) depths. The ΔL values at 0.7 mm were significantly lower than those at 1.5 mm (p = 0.004). The ΔE00 values at 0.7 mm were significantly lower than those at depths of 1.0 mm (p = 0.005) and 1.5 mm (p = 0.001). Except for the ΔL values at depths of 0.7 and 1.0 mm, statistically significant differences were found in ΔL and ΔE00 values in the pairwise comparisons of other depths. The ΔL and ΔE00 values at the dentin level were statistically higher than those at the enamel level. Further, no statistically significant difference was noted in Δa (p:0.148) and Δb (p: 0,471;) values among the depths.
Machine learning models
The R² values for all models were close to zero or negative, indicating that the depth variable alone was insufficient to explain the colour parameters. Machine learning models’ predictive success data in Fig. 1. Nevertheless, slight differences were observed among the models:
LR achieved the lowest MAE (3.4807), PR had the lowest MSE (24.6344) and the highest (albeit negative) R² value (–0.0009), Decision tree, random forest, and SVR models were at a similar level, with relatively low predictive performance.
- LR achieved the lowest MAE (3.4807).
- PR had the lowest MSE (24.6344) and the highest (albeit negative) R² value (–0.0009).
- Decision tree, random forest, and SVR models were at a similar level, with relatively low predictive performance.
These results showed that the overall predictive power of all models was limited, but the PR model demonstrated the best relative performance.
Discussion
This study examined the ΔE00, ΔL, Δa, and Δb values resulting from the preparation of anterior teeth at depths of 0.3, 0.5, 0.7, 1.0, and 1.5 mm, and assessed the performance of machine learning models in predicting colour components at different preparation depths based on the initial L*, a*, and b* values. In this context, it analyzed the predictive success of the models by comparing various algorithms such as LR, PR, DTR, RFR, and SVR. The Δa and Δb values did not change with depth, whereas the ΔL value changed. Therefore, the first H0, “colour values do not change with depth,” was partially rejected. Given the limited overall predictive power of all AI models, the second H0, “AI models are successful in predicting the abutment colour after preparation, before the preparation is performed,” was rejected.
When the ΔL, Δa, Δb, and ΔE00 values were evaluated for all teeth included in the study based on the preparation depth, the ΔL value changed significantly with an increase in the amount of preparation (Table1). The increase in ΔL values in deeper reduction (0.7–1.5 mm) compared with 0.3 mm and 0.5 mm preparations could be explained by the fact that the enamel tissue became thinner, which made the dentin-derived colour reflection more prominent on the surface. Ryakhovsky and Tikhon reported that the L value increased significantly with a decrease in enamel thickness in stepwise reduction of 0.3–0.6–0.9 mm [26]. Oguro et al., who evaluated reduction limited to the enamel level, reported that when enamel was thinned to approximately 100 µm, enamel thickness did not result in a significant difference in ΔL [22]. Similarly, the present study found that the measurements taken only at the enamel level did not yield a significant difference in ΔL at preparation depths of 0.3 and 0.5 mm. Additionally, at a preparation depth of 0.7 mm, dentin was visible on the measured surface in some samples, whereas at preparation depths of 1 and 1.5 mm, measurements were taken exclusively from the dentin surface.
No statistically significant difference was found in ∆a values (red–green axis) across all preparation depths. ∆a values did not increase in this study due to the use of extracted teeth, and no significant difference was found, consistent with studies demonstrating that this red visibility disappeared in teeth with the pulp removed [27, 28].
Studies have shown that when the b* value is evaluated separately for enamel and dentin, dentin has a more yellow b* value of 15–25 [21, 29], whereas enamel yields more neutral b* values compared with dentin [21, 29]. Pustina-Krasniqi et al. [21] found that when enamel and dentin were measured together, the average b* value was 21.6. Considering that dentin plays a more decisive role in determining tooth colour, the changes in the b* value as one approaches the dentin do not result in a statistically significant difference. Similarly, in this study, the changes in ∆b (yellow–blue axis) values were independent of preparation depth when all teeth were evaluated, implying no significant difference (Table1) [26]. Ryakhovsky et al., in line with our study, stated: “It is found that hue values depend only slightly on the depth and location of the preparation, with the tooth reflecting a relatively constant range of visible wavelengths”
No significant difference in ΔE00 values was observed between preparation depths of 0.3 and 0.5 mm, 0.5 and 0.7 mm, and 1.0 and 1.5 mm (Table 1). Also, a significant difference was observed at the boundary where the enamel–dentin junction was present [22]. No statistically significant difference was found in ∆E values between the 1.0 mm and 1.5 mm preparations in this study. This suggested that the colour might remain relatively stable in the dentin tissue following the removal of the enamel layer, regardless of whether the preparation depth was 1.0 or 1.5 mm.
In this study, five different machine learning algorithms (LR, PR, DTR, RFR, and SVR) were selected to predict the L*, a*, and b* components of tooth colour based on preparation depth. This was mainly to test both linear and nonlinear possibilities of the depth–colour change relationship, determine which type of mathematical model was more suitable for this relationship, and reveal the most clinically predictable approach.
It was assumed in this study that a linear change might occur in the L*, a*, and b* values with an increase in depth. Therefore, LR was used as the baseline model to assess whether the relationship between the depth variable and colour change exhibited a linear trend. However, LR produced low R² values and revealed that the depth–colour relationship was not linear. Therefore, a PR model was also applied to test the possible curvilinear nature of this relationship. However, both the LR and PR models yielded low R² values. This indicated that depth alone was insufficient to predict colour components. Tejada-Casado et al. (2023) [6] prepared a total of 15 different shades of VITAPAN Excell, VITAPAN Dentine and VITA Physiodens as well as VITA Enamel of five different thicknesses (0.5–2.5 mm range) were used to manufacture monolithic and layered samples. Also, they evaluated the relationship between thickness and CIE-Lab* colour coordinates using PR models and established a strong and predictable relationship between thickness and CIE-Lab* values [6]. In the present study, PR models demonstrated not high accuracy rates at low ΔE00 values. The failure of regression models in natural teeth may be attributed to the fact that the structure of enamel and dentin is much more heterogeneous compared with fabricated materials. Enamel thickness varies significantly among individuals and tooth types, whereas the colour components of the dentin surface vary widely depending on biological and anatomical variations [6, 30,31,32]. Moreover, the enamel–dentin ratio varied at each preparation stage between depths of 0.3 to 1.5 mm, making the optical behavior extremely complex to be adequately explained by a linear relationship. However, the layers were homogeneous, the colour distribution was uniform, and the optical properties were controlled in the composite samples used by Tejada-Casado et al. Therefore, thickness-dependent changes were more predictable, and regression models yielded successful results [6].
Kose et al. (2023) [11] evaluated 28 different regression models using a multidimensional dataset comprising various substrate colours, ceramic shades, and thicknesses. They reported that DTR was the most successful model among them, with the lowest MAE and the highest R² values [11]. The success of the decision could be attributed to the high variance in the dataset’s variables—that is, the wide distribution of data—the use of a broad set of independent variables (infrastructure and ceramic L*, a*, and b* coordinates, the use of ceramics with two different translucency values, and different ceramic thicknesses), and a large and balanced test set of 634 samples. The low predictive success of DTR in the present study could primarily be explained by fundamental differences in the data structure. Unlike the study by Kose et al. [11], this study used only preparation depth as an independent variable; other important factors influencing colour change (individual variations in enamel thickness, dentin–enamel ratio, surface gloss, optical path length, dentin-canal anatomy, etc.) were not included in the model. In other words, distinct subgroups sufficient to form the branching structure of the decision tree model did not emerge, and therefore, the model could not account for high variance, resulting in R² values that were close to zero or negative.
Previous studies have shown that the random forest algorithm performs well in modeling complex relationships in the dental field [12, 13]. This algorithm is capable of modeling nonlinear relationships in data with a heterogeneous structure [8, 33]. For this reason, random forest was selected as a suitable algorithm to capture the nonlinear nature of the depth–colour relationship in the present study and provide a complementary approach to other models. Additionally, other algorithms exhibited low prediction performance when estimating L*, a*, and b* values based on the preparation depth. Therefore, SVR was considered significantly advantageous given the sample size in this study because it demonstrated superior generalization performance on small and medium-sized datasets [34]. Furthermore, these algorithms failed to demonstrate the expected high estimation performance in this study. This was because the depth variable could not explain colour on its own, biological variation was high, and individual differences existed in enamel–dentin ratios [34, 35].
The findings of this study showed that the relationship between preparation depth and changes in colour components exhibited a biologically meaningful, yet mathematically unpredictable, level of complexity. In particular, the optical properties of the enamel layer, varying enamel thicknesses among individuals and tooth types, the high chromatic effect of dentin, and the individual-specific characteristics of the enamel–dentin junction made the relationship between L*, a*, and b*values and depth extremely multidimensional to be reduced to a linear or polynomial model. Therefore, the R² values of all regression models used in this study were close to 0 or a negative value because of the extremely heterogeneous structure of natural dental tissues rather than model failure. The only independent variable in this study was preparation depth. Therefore, the complex optical contributions of dentin and enamel could not be explained by a single parameter, and the models could not produce meaningful predictive power. Consequently, the poor performance of machine learning analyses likely reflected not algorithmic inadequacy, but the multidimensional, irregular, and individual-specific optical response of natural dental tissues to changes in thickness.
The study had some limitations. The teeth used in the study were devitalized, thereby eliminating the optical effect that pulp tissue contributes to tooth colour. Additionally, the amounts of enamel and dentin in the teeth were not standardized because the teeth were randomly selected.
Conclusions
1. The dentin effect became more pronounced, and the ΔL and ΔE00 values increased with an increase in the preparation depth in the anterior teeth. The optical effect of dentin became dominant with an increase in the amount of preparation. When Δa and Δb values were assessed for all teeth, they were found to be independent of preparation depth.
2. All machine learning models (LR, PR, DTR, RFR, and SVR) produced low R² values. Preparation depth alone was insufficient for predicting natural tooth colour.
Ethical declaratıons
There is no funding from an organisation or company directly for the research. None of the authors hold any advisory or consultative role with an organisation or company involved in this research or similar research. None of the authors hold any patents, shareholdings or board memberships in any company, nor have they received any speaking fees from any company.
Data availability
The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.
References
Koumpia E, Athanasiou AE, Eliades T, Knösel M. Precision of a reflectance spectrophotometer in measuring anterior tooth color. Open Dent J. 2018;12:884–95.
Dias R, Alves CP, Yehudi R, Guerra F, Messias A. Colorimetric behaviour of ceramic zirconia restorations cemented on darkened substrates—in vitro study. Surfaces. 2026;9:27.
Chen D, He X, Li Q, Wang Z, Shen J, Shen J. Predicting enamel depth distribution of maxillary teeth based on intraoral scanning: a machine learning study. J Prosthodontic Res. 2026;70:73–81.
Ardila CM, Pulgarín-Medina DM, Pineda-Vélez E, Vivares-Builes AM. Artificial intelligence for color prediction and esthetic design in CAD/CAM Ceramic restorations: a systematic review and meta-analyses. Prosthesis. 2025;7:160.
Yılmaz S, Yılmaz HI, Jamshidi N. Emerging artificial intelligence technologies for evaluation of dental composite restorations: a scoping review. Eur J Dentistry. 2026;
Tejada-Casado M, Ghinea R, Pérez MM, Ruiz-López J, Lübbe H, Herrera LJ. Development of thickness-dependent predictive methods for the estimation of the CIEL* a* b* color coordinates of monolithic and layered dental resin composites. Materials. 2023;16:761.
Arjumand B. The application of artificial intelligence in restorative dentistry: a narrative review of current research. Saudi Dent J. 2024;36:835–40.
Sharma A. Advancements in dental science through artificial intelligence and machine learning: a comprehensive review. Int J Clin Pediatr Dent. 2026;19:120–6.
Alzubi J, Nayyar A, Kumar A Machine learning from theory to algorithms: an overview. IOP Publishing; 2018. p. 012012.
Dave M Ten years of BDJ Open: reflections on a decade of transformation in dentistry: Ten years of BDJ Open: reflections on a decade of transformation in dentistry. Nature Publishing Group UK London; 2025.
Kose C Jr, Oliveira D, Pereira PN, Rocha MG. Using artificial intelligence to predict the final color of leucite-reinforced ceramic restorations. J Esthet Restor Dent. 2023;35:105–15.
Wang R, Li K, Su Y-X. Prediction of ameloblastoma recurrence using random forest—a machine learning algorithm. Int J Oral Maxillofac Surg. 2022;51:886–91.
Regueira-Iglesias A, Suárez-Rodríguez B, Blanco-Pintos T, et al. Diversity and random forest models of oral microbiomes in periodontal health using publicly available data. J Periodontol. 2025.
Cherkassky V, Ma Y Selection of meta-parameters for support vector regression. Springer; 2002:687–93.
Nichols JA, Herbert Chan HW, Baker MA. Machine learning: applications of artificial intelligence to imaging and diagnosis. Biophys Rev. 2019;11:111–8.
Vafaei F, Izadi A, Abbasi S, Farhadian M, Bagheri Z. Comparison of optical properties of laminate veneers made of Zolid FX and Katana UTML zirconia and lithium disilicate ceramics. Front Dent. 2019;16:357.
Cattoni F, Teté G, Calloni AM, Manazza F, Gastaldi G, Capparè P. Milled versus moulded mock-ups based on the superimposition of 3D meshes from digital oral impressions: a comparative in vitro study in the aesthetic area. BMC Oral Health. 2019;19:230.
Almedaires M, Delgado A, Abdulhameed N, Pereira P, Rocha MG, Oliveira D. Predicting final restoration color using neural network models: the impact of substrate lightness versus ceramic shade, translucency and thickness. J Esthet Restor Dent. 2025;37:1900–6.
Asiri SA, Jameel TS, Eyad M, Ba-shammakh S, Alghamdi HS, Ashy LM. Evaluating the predictive performance of newly developed machine learning models for shade selection in zirconia and lithium disilicate ceramics: an in vitro study. Int J Dent. 2026;2026:1628893.
Abichandani SJ, Dutta A. Predictors of patient dissatisfaction with ceramic veneers: a matched case–control study. J Indian Prosthodontic Soc. 2026;26:96–101.
Pustina-Krasniqi T, Shala K, Staka G, Bicaj T, Ahmedi E. Dula L. Lightness, chroma, and hue distributions in natural teeth measured by a spectrophotometer. Eur J Dent. 2017;11:036–40.
Oguro R, Nakajima M, Seki N, Sadr A, Tagami J, Sumi Y. The role of enamel thickness and refractive index on human tooth colour. J Dent. 2016;51:36–44.
Rajput D, Wang W-J, Chun-Chuan C. Evaluation of a decided sample size in machine learning applications. BMC Bioinforma. 2023;24:1.
Wanna Y, Wiratchawa K, Leenaracharoongruang R, Sittiwong W, Panpisut P, Intharah T Dentshadeai: a framework for automatic dental shade matching through mobile phone camera. IEEE; 2022:282–5.
Kim M, Kim B, Park B, et al. A digital shade-matching device for dental color determination using the support vector machine algorithm. Sensors. 2018;18:3051.
Ryakhovsky A, Tikhon YV. Analysis of the color changes in teeth at different depths preparation. Stomatologiia. 2017;96:40–43.
Ardu S, Feilzer AJ, Devigus A, Krejci I. Quantitative clinical evaluation of esthetic properties of incisors. Dent Mater. 2008;24:333–40.
Turgut S, Kilinç H, Eyüpoglu GB, Bağis B. Color relationships of natural anterior teeth: an: in vivo: study. Niger J Clin Pract. 2018;21:925–31.
Pop-Ciutrila I-S, Ghinea R, Colosi HA, et al. Color compatibility between dental structures and three different types of ceramic systems. BMC Oral Health. 2021;21:75.
Gómez-Polo C, Montero J, Gómez-Polo M, de Parga JAMV, Celemin-Viñuela A. Natural tooth color estimation based on age and gender. J Prosthodontics. 2017;26:107–14.
Al-Zahawi AR, Rawa OI, Talabani RM, Dawood SN, Garib DSH, Abdalla AO. Age and sex related change in tooth enamel thickness of maxillary incisors measured by cone beam computed tomography. BMC Oral Health. 2023;23:1.
Liondev I, Dzhondrova I, Radev R, Kirov D, Radeva E, Todorov R. Color characteristics of the anterior teeth in dental students. J Imab. 2025;31:3.
Carrillo-Perez F, Pecho OE, Morales JC, et al. Applications of artificial intelligence in dentistry: a comprehensive review. J Esthet Restor Dent. 2022;34:259–80.
Lahmiri S, Bekiros S, Bezzina F. Complexity analysis and forecasting of variations in cryptocurrency trading volume with support vector regression tuned by Bayesian optimization under different kernels: an empirical comparison from a large dataset. Expert Syst Appl. 2022;209:118349.
Hein S, Modrić D, Westland S, Tomeček M. Objective shade matching, communication, and reproduction by combining dental photography and numeric shade quantification. J Esthet Restor Dent. 2021;33:107–17.
Acknowledgements
We would like to thank Associate Professor Ayşegül Kurt, Head of the Department of Prosthetic Dentistry at Trakya University Faculty of Dentistry, for her support.
Author information
Authors and Affiliations
Department of Prosthodontics, Dentistry Faculty, Trakya University, Edirne, Turkey
Tugba Kuyumcu
Department of Prosthodontics, Faculty of Dentistry, Trakya University, Edirne, Turkey
Onur Dogan DAG
Authors
- Tugba Kuyumcu
- Onur Dogan DAG
Contributions
Both authors have—made significant contributions to the concept or design of the study; or to the collection, analysis or interpretation of the data; and—have drafted the manuscript or critically reviewed it for its intellectual content; and—have approved the version to be published; and—have contributed to ensuring that any questions regarding the accuracy or integrity of any part of the work are properly investigated and resolved, and have accepted responsibility for all aspects of the work.
Corresponding author
Correspondence to Onur Dogan DAG.
Ethics declarations
Competing interests
The authors declare no competing interests.
Informed consent
Informed consent has been obtained from all participants in the study. It was approved by the Ethics Committee for Non-Interventional Scientific Research at the Faculty of Medicine, Trakya University (date: January 6, 2025; approval number 01/27).
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Kuyumcu, T., DAG, O.D. Evaluation of colour parameters at different preparation depths and prediction of tooth colour using machine learning. BDJ Open 12, 109 (2026). https://doi.org/10.1038/s41405-026-00508-y
Received: 27 May 2026
Revised: 28 September 2026
Accepted: 07 October 2026
Published: 07 November 2026
Version of record: 07 November 2026
DOI: https://doi.org/10.1038/s41405-026-00508-y
来源:Nature 健康护理主题 · nature.com