Weakly Supervised Histopathology Image Segmentation With Sparse Point Annotations.

Area of Science:

• Digital pathology
• Computer-assisted diagnostics
• Machine learning for medical imaging

Background:

• Digital histopathology image segmentation aids cancer diagnostics.
• Manual annotations are difficult, making weak supervision preferable.
• Existing weakly supervised models struggle with histopathology image heterogeneity.

Purpose of the Study:

• Introduce WESUP, a novel end-to-end weakly supervised learning framework.
• Achieve accurate segmentation and generalizability using sparse point annotations.
• Address limitations of current weakly supervised methods in histopathology.

Main Methods:

• Employ hierarchical feature representation using superpixels for local and global context.
• Utilize deep dynamic label propagation to identify targets from sparse annotations.
• Jointly train both components and use pixel-wise inference for enhanced prediction.

Main Results:

• WESUP effectively distinguishes histological foreground from background.
• Outperforms state-of-the-art weakly supervised methods on diverse histopathology datasets.
• Achieves superior performance compared to fully supervised networks with minimal annotations.

Conclusions:

• WESUP offers a powerful solution for weakly supervised histopathology image segmentation.
• Demonstrates the potential of minimal annotation strategies in medical imaging.
• Advances computer-assisted cancer diagnostics through improved segmentation accuracy.