Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Segmentation Style Discovery: Application to Skin Lesion Images

Kumar Abhishek, Jeremy Kawahara, Ghassan Hamarneh

TL;DR

The paper tackles segmentation style discovery for medical images without annotator correspondence, introducing StyleSeg, a joint model that yields $M$ plausible, diverse, and semantically consistent segmentation styles from image–mask corpora. It optimizes a segmentation model $f_s$ and a style classifier $f_c$ using losses $\,\mathcal{L}_1$, $\mathcal{L}_2$, and $\mathcal{L}_3$ to align predictions with ground-truth via Dice while maintaining style plausibility, and introduces AS2 to quantify how well styles align with annotator preferences. The authors validate StyleSeg on four skin lesion segmentation datasets, showing improvements over single- and multi-hypothesis baselines, and curate ISIC-MultiAnnot (12{,}951 images, 10 annotators, 13{,}555 image–mask pairs) to demonstrate strong style–annotator alignment across 27 preferences. The work enables annotation-style-aware segmentation without explicit annotator IDs, with practical impact for personalized clinical analysis and dataset curation, and points to future work on disentangling content from style and optimizing the number of styles.

Abstract

Variability in medical image segmentation, arising from annotator preferences, expertise, and their choice of tools, has been well documented. While the majority of multi-annotator segmentation approaches focus on modeling annotator-specific preferences, they require annotator-segmentation correspondence. In this work, we introduce the problem of segmentation style discovery, and propose StyleSeg, a segmentation method that learns plausible, diverse, and semantically consistent segmentation styles from a corpus of image-mask pairs without any knowledge of annotator correspondence. StyleSeg consistently outperforms competing methods on four publicly available skin lesion segmentation (SLS) datasets. We also curate ISIC-MultiAnnot, the largest multi-annotator SLS dataset with annotator correspondence, and our results show a strong alignment, using our newly proposed measure AS2, between the predicted styles and annotator preferences. The code and the dataset are available at https://github.com/sfu-mial/StyleSeg.

Segmentation Style Discovery: Application to Skin Lesion Images

TL;DR

The paper tackles segmentation style discovery for medical images without annotator correspondence, introducing StyleSeg, a joint model that yields plausible, diverse, and semantically consistent segmentation styles from image–mask corpora. It optimizes a segmentation model and a style classifier using losses , , and to align predictions with ground-truth via Dice while maintaining style plausibility, and introduces AS2 to quantify how well styles align with annotator preferences. The authors validate StyleSeg on four skin lesion segmentation datasets, showing improvements over single- and multi-hypothesis baselines, and curate ISIC-MultiAnnot (12{,}951 images, 10 annotators, 13{,}555 image–mask pairs) to demonstrate strong style–annotator alignment across 27 preferences. The work enables annotation-style-aware segmentation without explicit annotator IDs, with practical impact for personalized clinical analysis and dataset curation, and points to future work on disentangling content from style and optimizing the number of styles.

Abstract

Variability in medical image segmentation, arising from annotator preferences, expertise, and their choice of tools, has been well documented. While the majority of multi-annotator segmentation approaches focus on modeling annotator-specific preferences, they require annotator-segmentation correspondence. In this work, we introduce the problem of segmentation style discovery, and propose StyleSeg, a segmentation method that learns plausible, diverse, and semantically consistent segmentation styles from a corpus of image-mask pairs without any knowledge of annotator correspondence. StyleSeg consistently outperforms competing methods on four publicly available skin lesion segmentation (SLS) datasets. We also curate ISIC-MultiAnnot, the largest multi-annotator SLS dataset with annotator correspondence, and our results show a strong alignment, using our newly proposed measure AS2, between the predicted styles and annotator preferences. The code and the dataset are available at https://github.com/sfu-mial/StyleSeg.
Paper Structure (6 sections, 8 equations, 4 figures, 2 tables)

This paper contains 6 sections, 8 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Results and Discussion
  4. Conclusion
  5. Acknowledgments.
  6. Disclosure of Interests.

Figures (4)

  • Figure 1: (a) An overview of the proposed method StyleSeg. (b) Inter-annotator variability in the training images. (c) An annotator-wise breakdown of the newly curated ISIC-MultiAnnot dataset.
  • Figure 2: Evaluating StyleSeg on ISIC Archive-Test: diverse and plausible segmentations that are semantically consistent across styles.
  • Figure SM1: Supplementary Figures
  • Figure SM2: (continued) Supplementary Figures.