Weakly supervised segmentation on neural compressed histopathology with self-equivariant regularization.

Area of Science:

• Digital pathology
• Computational biology
• Medical image analysis

Background:

• Accurate segmentation of histology images is crucial for disease diagnosis and treatment.
• Fully supervised methods require laborious pixel-level annotations, while patch-based approaches can lose global context.
• Weakly supervised learning using image-level labels offers a more practical alternative.

Purpose of the Study:

• To develop a weakly supervised framework for histopathology segmentation using only image-level labels.
• To address the limitations of existing fully supervised and patch-based segmentation methods.
• To improve the efficiency and practicality of training segmentation algorithms in digital pathology.

Main Methods:

• A weakly supervised framework was developed, compressing gigapixel histology images using unsupervised contrastive learning to preserve spatial context.
• A network was trained on compressed images to predict image-labels and refine initial class activation maps (CAMs) via self-supervised losses.
• Refinement was achieved using a pixel correlation module (PCM) with self-attention and a feature masking technique for spatial dropout.

Main Results:

• The proposed method achieved segmentation performance comparable to fully supervised approaches.
• The framework outperformed existing state-of-the-art patch-based methods on two curated datasets.
• The integration of CAM refinement with self-supervision demonstrated effectiveness in histopathology segmentation.

Conclusions:

• Weakly supervised learning with refined CAMs and self-supervision is a viable and effective approach for histopathology segmentation.
• The developed framework offers a practical alternative to laborious annotation-intensive methods.
• This work advances automated analysis in digital pathology by leveraging readily available image-level labels.