Artifact-aware Lightweight U-Net用于皮肤镜病灶分割及抗干扰鲁棒性评测
Artifact-aware lightweight dermoscopic lesion segmentation with reproducible corruption robustness evaluation
该研究提出一种抗伪影轻量级U形网络ARL-UNet,通过多线索主干估计毛发、反光、气泡和皮纹等采集伪影区域,并在有限参数预算下保留病灶特征;基于ISIC 2017/2018构建含四种临床相关伪影及两个严重程度的可复现 corrupt benchmark 进行评估,课程训练策略在保持清洁域精度的同时提升了污损域的鲁棒性,并在PH2数据集上实现零微调迁移。
Abstract
Automated dermoscopic lesion segmentation is a key step in computer-aided skin cancer analysis, yet lightweight models for resource-constrained dermoscopic image analysis can be unstable when images contain acquisition artifacts such as hair, specular highlights, skin lines, and air bubbles. This paper addresses this problem by proposing an artifact-aware lightweight U-shaped segmentation framework together with a reproducible corruption benchmark for robustness evaluation. The proposed framework first estimates artifact-prone regions through a multi-cue stem that combines Laplacian, bright-region, dark-region, and learned responses. It then uses efficient multi-scale encoder blocks with depthwise separable operations and channel gating to preserve lesion-relevant features under a limited parameter budget. During decoding, a mask-guided skip-fusion module integrates coarse lesion predictions with propagated artifact cues, reducing the direct transfer of contaminated shallow features while retaining boundary information. To evaluate robustness beyond clean ISIC images, we construct fixed corrupted subsets from ISIC 2017 and ISIC 2018 by applying four clinically motivated artifact categories at two severity levels, with unchanged ground-truth masks and recorded artifact metadata. Experiments compare clean-only training, direct artifact augmentation, and curriculum-based artifact exposure against representative lightweight baselines. The results show that the proposed framework maintains competitive clean-domain accuracy, improves corrupted-domain robustness, benefits from curriculum training under stronger artifact severity, and transfers better to PH2 without fine-tuning. Ablation studies further verify the complementary effects of artifact-aware input modeling, robust encoder design, and mask-guided skip fusion. Overall, this work provides a lightweight segmentation framework and a reproducible evaluation protocol for studying artifact robustness in dermoscopic image segmentation. Our code is available at: https://github.com/28001736/ARL-UNet.
Funding
This work was supported by the Guangxi Science and Technology Program under Grant Nos. GuiKeFN2600640336 and GuiKeFN2600640049, and the Natural Science Foundation of Guangxi under Grant No. 2026GXNSFAA00640439.
Author information
Authors and Affiliations
Guangxi Key Laboratory of Machine Vision and Intelligent Control, Wuzhou University, Wuzhou, 543002, China
Haixiao Gong, Jie He, Hui Guo & Linting Meng
School of Electronic Information and Artificial Intelligence, Wuzhou University, Wuzhou, 543000, China
Haixiao Gong, Jie He & Linting Meng
Wuzhou Vocational College, Wuzhou, 543002, China
Cuijiao Yin & Mei Tang
School of Tourism, Qinghai Minzu University, Xining, 810007, China
Chengchen Gong
Authors
- Haixiao Gong
- Cuijiao Yin
- Jie He
- Hui Guo
- Mei Tang
- Linting Meng
- Chengchen Gong
Corresponding author
Correspondence to Cuijiao Yin.
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
A. Detailed results of corrupted dermoscopic benchmarks
This section provides the full numerical results corresponding to the robustness analyses summarized in the main text of "Robustness on corrupted dermoscopic benchmarks". While the main paper reports the key trends using compact visual summaries, the tables in this section list the complete evaluation metrics for each dataset, training regime, corruption severity, and artifact type.
Tables 12 and 13 report the main robustness comparison on the corrupted ISIC 2017 and ISIC 2018 benchmarks, respectively. For each training strategy \(\mathcal {T}_0\)–\(\mathcal {T}_2\), we provide results under both Severity 1 and Severity 2 corruptions. These tables include mIoU, Dice, Accuracy, Specificity, Sensitivity, and HD95, allowing a more complete comparison of region overlap, pixel-level classification, and boundary localization performance.
Tables 14 and 15 further present the per-artifact mIoU breakdown for Terminal Hair, Specular Highlight, Air Bubble, and Skin Line corruptions. These results complement the main analysis by showing how each model behaves under different visual disturbance patterns. Overall, the detailed tables confirm the trends discussed in Section "Robustness on corrupted dermoscopic benchmarks": ARL-UNet consistently achieves stronger corrupted-domain robustness, particularly under severe corruptions and artifact types that interfere with lesion boundaries.
B. Detailed results of PH2 zero-shot evaluation
This section provides the full numerical results for the PH2 zero-shot evaluation discussed in the main text. The main paper presents the results as compact visual summaries, while the tables in this section report the complete metric values for each source dataset and training regime.
Tables 16 and 17 show the zero-shot performance on PH2 when models are trained on ISIC 2017 and ISIC 2018, respectively. For each training strategy \(\mathcal {T}_0\)–\(\mathcal {T}_2\), we report mIoU, Dice, Accuracy, Specificity, Sensitivity, and HD95. These metrics provide a detailed comparison of region overlap, pixel-level classification, and boundary localization under cross-dataset transfer.
The detailed results support the observations in Section "PH2 Zero-shot evaluation". ARL-UNet consistently achieves stronger performance in the principal segmentation metrics, especially mIoU, Dice, and HD95, across both ISIC source domains. The tables also show that curriculum learning \(\mathcal {T}_2\) provides the best overall transfer performance for ARL-UNet, confirming that artifact-aware training improves robustness without degrading zero-shot generalization to PH2.
C. Visualization results
We provide additional qualitative results to illustrate model behavior under different dermoscopic artifacts. Each figure compares predictions on the clean image and its corrupted variants, including hair, highlight, air bubble, and skin-line artifacts at two severity levels.
Qualitative comparison under different artifact corruptions.
As shown in Fig. 11, artifact corruptions introduce different failure patterns. UCM-Net often produces smaller masks and misses peripheral lesion regions, while EMA-UNet tends to generate over-smoothed predictions. U-Net preserves larger regions in some cases, but its masks are less stable under severe highlights, bubbles, and skin-line artifacts.
In contrast, ARL-UNet better maintains the overall lesion shape across corruption types. It preserves the elongated lesion structure more reliably, especially when hair or skin-line artifacts overlap the lesion. These visual results are consistent with the quantitative findings, showing that ARL-UNet reduces artifact-induced mask distortion while retaining lesion boundaries.
D. Curriculum training settings for ablation
This section provides the detailed curriculum schedules used in the curriculum-training ablation study. In the main curriculum setting, the model is first trained on clean images and then exposed to artifact-corrupted samples with gradually increased corruption probability and severity. To analyze the contribution of different curriculum components, we consider two variants: curriculum training without clean warm-up and curriculum training without Severity-2 corruptions.
Table 18 shows the schedule for the curriculum variant without clean warm-up. In this setting, artifact-corrupted samples are introduced from the beginning of training. The first stage uses only mild corruptions with a sampling probability of 0.50, while the second stage increases the probability to 0.75 and includes both Severity 1 and Severity 2 corruptions. This variant is used to test whether early exposure to artifacts affects the learning of stable clean-domain lesion representations. Here, \(\mathcal {A}\)={Hair, highlight, air bubble, skin line}.
Table 19 presents the curriculum variant without Severity 2 corruptions. This setting keeps the clean warm-up stage but introduces only mild artifact corruptions afterward. It is designed to examine whether exposure to severe corruptions is necessary for improving robustness under stronger artifact perturbations.
Together, these schedules isolate two key factors in curriculum design: the clean warm-up phase and the inclusion of severe corruptions. The corresponding ablation results in the main text show that both components are important for achieving a favorable balance between clean-domain accuracy and corrupted-domain robustness.
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
Gong, H., Yin, C., He, J. et al. Artifact-aware lightweight dermoscopic lesion segmentation with reproducible corruption robustness evaluation. Sci Rep (2026). https://doi.org/10.1038/s41598-026-74096-0
Received: 30 June 2026
Accepted: 28 September 2026
Published: 14 November 2026
DOI: https://doi.org/10.1038/s41598-026-74096-0
Keywords
来源:Nature 健康护理主题 · nature.com