A Study of Artifacts on Melanoma Classification under Diffusion-Based Perturbations

Qixuan Jin, Marzyeh Ghassemi
Proceedings of the sixth Conference on Health, Inference, and Learning, PMLR 287:844-861, 2025.

Abstract

In melanoma classification, deep learning models have been shown to rely on non-medical artifacts (e.g., surgical markings) rather than clinically relevant features (e.g., lesion asymmetry), compromising their generalizability. In this work, we investigate the impact of artifacts on melanoma classification under two settings: (1) input disruptions, such as bounding boxes and frequency-based filtering, which isolate artifacts by region or frequency, and (2) a novel diffusion-based perturbation method that selectively introduces isolated artifacts into images, generating controlled pairs for direct comparison. We systematically analyze artifact biases in three benchmark datasets: ISIC 2018, HAM10000, and PH2. Our findings reveal that whole-image training outperforms lesion-only or background-only approaches, low-frequency features are essential for melanoma prediction, and classifiers are more sensitive to perturbations for the artifacts of ink markings, rulers, and patches. These results emphasize the need for systematic artifact assessment and provide insights for improving the robustness of melanoma classification models.

Cite this Paper

BibTeX
@InProceedings{pmlr-v287-jin25b, title = {A Study of Artifacts on Melanoma Classification under Diffusion-Based Perturbations}, author = {Jin, Qixuan and Ghassemi, Marzyeh}, booktitle = {Proceedings of the sixth Conference on Health, Inference, and Learning}, pages = {844--861}, year = {2025}, editor = {Xu, Xuhai Orson and Choi, Edward and Singhal, Pankhuri and Gerych, Walter and Tang, Shengpu and Agrawal, Monica and Subbaswamy, Adarsh and Sizikova, Elena and Dunn, Jessilyn and Daneshjou, Roxana and Sarker, Tasmie and McDermott, Matthew and Chen, Irene}, volume = {287}, series = {Proceedings of Machine Learning Research}, month = {25--27 Jun}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v287/main/assets/jin25b/jin25b.pdf}, url = {https://proceedings.mlr.press/v287/jin25b.html}, abstract = {In melanoma classification, deep learning models have been shown to rely on non-medical artifacts (e.g., surgical markings) rather than clinically relevant features (e.g., lesion asymmetry), compromising their generalizability. In this work, we investigate the impact of artifacts on melanoma classification under two settings: (1) input disruptions, such as bounding boxes and frequency-based filtering, which isolate artifacts by region or frequency, and (2) a novel diffusion-based perturbation method that selectively introduces isolated artifacts into images, generating controlled pairs for direct comparison. We systematically analyze artifact biases in three benchmark datasets: ISIC 2018, HAM10000, and PH2. Our findings reveal that whole-image training outperforms lesion-only or background-only approaches, low-frequency features are essential for melanoma prediction, and classifiers are more sensitive to perturbations for the artifacts of ink markings, rulers, and patches. These results emphasize the need for systematic artifact assessment and provide insights for improving the robustness of melanoma classification models.} }
Endnote
%0 Conference Paper %T A Study of Artifacts on Melanoma Classification under Diffusion-Based Perturbations %A Qixuan Jin %A Marzyeh Ghassemi %B Proceedings of the sixth Conference on Health, Inference, and Learning %C Proceedings of Machine Learning Research %D 2025 %E Xuhai Orson Xu %E Edward Choi %E Pankhuri Singhal %E Walter Gerych %E Shengpu Tang %E Monica Agrawal %E Adarsh Subbaswamy %E Elena Sizikova %E Jessilyn Dunn %E Roxana Daneshjou %E Tasmie Sarker %E Matthew McDermott %E Irene Chen %F pmlr-v287-jin25b %I PMLR %P 844--861 %U https://proceedings.mlr.press/v287/jin25b.html %V 287 %X In melanoma classification, deep learning models have been shown to rely on non-medical artifacts (e.g., surgical markings) rather than clinically relevant features (e.g., lesion asymmetry), compromising their generalizability. In this work, we investigate the impact of artifacts on melanoma classification under two settings: (1) input disruptions, such as bounding boxes and frequency-based filtering, which isolate artifacts by region or frequency, and (2) a novel diffusion-based perturbation method that selectively introduces isolated artifacts into images, generating controlled pairs for direct comparison. We systematically analyze artifact biases in three benchmark datasets: ISIC 2018, HAM10000, and PH2. Our findings reveal that whole-image training outperforms lesion-only or background-only approaches, low-frequency features are essential for melanoma prediction, and classifiers are more sensitive to perturbations for the artifacts of ink markings, rulers, and patches. These results emphasize the need for systematic artifact assessment and provide insights for improving the robustness of melanoma classification models.
APA
Jin, Q. & Ghassemi, M.. (2025). A Study of Artifacts on Melanoma Classification under Diffusion-Based Perturbations. Proceedings of the sixth Conference on Health, Inference, and Learning, in Proceedings of Machine Learning Research 287:844-861 Available from https://proceedings.mlr.press/v287/jin25b.html.

Related Material