Weakly supervised joint whole-slide segmentation and classification in prostate cancer.

Area of Science:

• Digital Pathology
• Computational Pathology
• Medical Image Analysis

Background:

• Histological region segmentation aids pathologists but requires costly pixel-level annotations for whole-slide images (WSIs).
• Weakly-supervised methods for WSI classification exist, but WSI-level segmentation with limited supervision is underexplored.
• Existing WSI segmentation methods often need supervision beyond image labels, limiting practical application.

Purpose of the Study:

• To propose WholeSIGHT, a novel weakly-supervised method for simultaneous segmentation and classification of WSIs.
• To address the challenge of limited annotations in WSI analysis.
• To develop a method applicable to WSIs of arbitrary shapes and sizes.

Main Methods:

• WholeSIGHT constructs a tissue-graph representation of WSIs.
• A graph classification head generates pseudo-labels from image-level labels via feature attribution.
• Pseudo-labels train a node classification head for segmentation, enabling simultaneous prediction during testing.

Main Results:

• WholeSIGHT achieved state-of-the-art weakly-supervised segmentation performance across three public prostate cancer WSI datasets.
• The method demonstrated comparable or superior classification performance against existing weakly-supervised WSI classification techniques.
• The generalization capability, uncertainty estimation, and model calibration of WholeSIGHT were also assessed.

Conclusions:

• WholeSIGHT effectively performs weakly-supervised segmentation and classification of WSIs.
• The method overcomes the limitations of expensive pixel-level annotations.
• WholeSIGHT offers a promising tool for enhancing diagnostic support in digital pathology.