Semi-Supervised Skin Lesion Segmentation under Dual Mask Ensemble with Feature Discrepancy Co-Training

Thanh-Huy Nguyen, Thien Nguyen, Xuan Bach Nguyen, Nguyen Lan Vi Vu, Vinh Quang Dinh, Fabrice MERIAUDEAU

Keywords: Semi-Supervised Learning, Skin cancer diagnosis, Co-training Network

TL;DR: A novel feature discrepancy loss trained with a cross-pseudo supervision strategy and Dual Mask Ensemble training that enables the sub-models to extract more meaningful information from unlabeled data by combining mask predictions.

Abstract: Skin Lesion Segmentation with supportive Deep Learning has become essential in skin lesion analysis and skin cancer diagnosis. However, in the practical scenario of clinical implementation, there is a limitation in human-annotated labels for training data, which leads to poor performance in supervised training models. In this paper, we propose Dual Mask Ensemble (DME) based on a dual-branch co-training network, which aims to enforce two models to exploit information from different views. Specifically, we introduce a novel feature discrepancy loss trained with a cross-pseudo supervision strategy, which enhances model representation by encouraging the sub-networks to learn from distinct features, thereby mitigating feature collapse. Additionally, Dual Mask Ensemble training enables the sub-models to extract more meaningful information from unlabeled data by combining mask predictions. Experimental results demonstrate the effectiveness of our approach, achieving state-of-the-art performance across several metrics (Dice and Jaccard) on the ISIC2018 and HAM10000 datasets. Our code is available at https://github.com/antares0811/DME-FD.

Primary Subject Area: Learning with Noisy Labels and Limited Data

Secondary Subject Area: Application: Dermatology

Paper Type: Methodological Development

Registration Requirement: Yes

Reproducibility: https://github.com/antares0811/DME-FD

Visa & Travel: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Created a single midl25_NNN.zip file with midl25_NNN.tex, midl25_NNN.bib, all necessary figures and files., Includes \documentclass{midl}, \jmlryear{2025}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package, Did not use the times package., All authors and co-authors are correctly listed with proper spelling and avoid Unicode characters., Author and institution details are de-anonymized where needed. All author names, affiliations, and paper title are correctly spelled and capitalized in the biography section., References must use the .bib file. Did not override the bibliographystyle defined in midl.cls. Did not use \begin{thebibliography} directly to insert references., Tables and figures do not overflow margins; avoid using \scalebox; used \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., Appendices and supplementary material are included in the same PDF after references., Main paper does not exceed 9 pages; acknowledgements, references, and appendix start on page 10 or later.

Submission Number: 94