WSSS-CRAM: precise segmentation of histopathological images via class region activation mapping.

Area of Science:

• Digital pathology
• Medical image analysis
• Computer vision

Background:

• Histopathological image analysis is crucial but often requires manual feature design or extensive labeled data for deep learning.
• Current deep learning methods for histopathological image analysis demand large amounts of labeled data, limiting their practical application.

Purpose of the Study:

• To develop a weakly-supervised semantic segmentation (WSSS) method for histopathological images.
• To enable accurate pixel-level annotation generation from limited image-level labels.
• To create an end-to-end trainable framework for automatic histopathological image analysis.

Main Methods:

• Introduced WSSS-CRAM, a novel weakly-supervised semantic segmentation framework.
• Utilized a discriminative activation strategy to generate category-specific activation maps from class labels.
• Employed conditional random fields for post-processing activation maps into reliable pseudo-ground-truth labels.
• Integrated pseudo-label acquisition and segmentation model training into an end-to-end joint training process.

Main Results:

• The WSSS-CRAM framework successfully predicts pixel-level labels using only image-level annotations.
• Demonstrated strong performance on quantitative evaluations and visualization results.
• The method performs well even when tested on images without any initial image-level annotations.

Conclusions:

• WSSS-CRAM effectively generates detailed pixel-level labels from coarse image-level annotations in histopathology.
• The proposed end-to-end framework offers a viable solution for automatic histopathological image analysis with reduced annotation burden.
• Future work will focus on validating the generalization capability across diverse pathological datasets and tissue image types.