By Youssef Karout (Primaa) and Nicolas Nerrienet (Primaa), Clara Simmat (Primaa), Stéphane Sockeel (Primaa), Rémy Peyret (Primaa)

Context

• Semantic segmentation of skin lesions in histopathology Whole Slide Images (WSIs) is hindered by the prohibitive cost of dense pixel-wise annotation and the pathological complexity of lesion boundaries.
• Significant inter-annotator variability introduces label noise, while the computational burden of processing giga-pixel images at high magnification limits clinical scalability.

Proposition

Our pipeline distills high-magnification expert knowledge into efficient low-magnification models for multiclass segmentation.

By curating high-quality labels from a multi-annotator subset, we propagate superior ground truth across the entire dataset.

Experiments & Results

Specalized models: HQ slides versus entire annotated dataset.

Improvement in F1-score achieved when training a model on HQ DS versus the models trained on all the annotated data.

Multiclass segmentation model.

Improvement in F1-score achieved when training a model only on annotated data versus the proposed pipeline.

Use case: Skin WSI classification

• Center A: 79% to 90% classification accuracy for all slides (189 WSIs).
• Center B: 85% to 90% accuracy for cancerous slides (89 WSIs) and 70% accuracy for healthy slides (21 WSIs).

Results

Conclusion

• Performance & Robustness: The model achieved high accuracy in multi-class skin lesion segmentation.
• Clinical Utility: It proved effective not just for segmentation, but also for slide-level classification tasks.