Weakly Supervised Gland Segmentation Based on Hierarchical Attention Fusion and Pixel Affinity Learning.

Area of Science:

• Digital pathology
• Medical image analysis
• Computational oncology

Background:

• Accurate gland segmentation in histopathology is crucial for colorectal cancer diagnosis.
• Current methods require labor-intensive pixel-level annotations.
• Weakly supervised learning offers a more efficient alternative.

Purpose of the Study:

• To develop an annotation-efficient, weakly supervised framework for gland segmentation.
• To improve the accuracy and robustness of gland segmentation in histopathological images.
• To reduce the cost and time associated with creating detailed annotations.

Main Methods:

• Proposed a two-stage weakly supervised segmentation framework: Multi-Level Attention and Affinity (MAA).
• Utilized image-level labels with Multi-Level Attention Fusion (MAF) for feature extraction and initial maps.
• Employed Affinity Refinement (AR) to refine pseudo-labels by modeling inter-pixel semantic consistency.

Main Results:

• The MAA framework achieved an Intersection over Union (IoU) of 81.99% and Dice coefficient of 90.10% on the GlaS dataset.
• Outperformed the state-of-the-art Online Easy Example Mining (OEEM) method by 4.43% in IoU.
• Demonstrated effective gland segmentation with reduced annotation effort.

Conclusions:

• Integrating hierarchical attention and affinity-guided refinement enhances annotation-efficient gland segmentation.
• The MAA framework provides a robust and accurate solution for histopathological image analysis.
• This approach holds significant potential for improving colorectal cancer diagnosis.