DIPathMamba: A domain-incremental weakly supervised state space model for pathology image segmentation.

Area of Science:

• Digital Pathology
• Medical Image Analysis
• Machine Learning

Background:

• Accurate pathology image segmentation is vital for digital pathology.
• Current methods struggle with dense annotations and handling diverse, large-scale datasets across multiple domains.
• Existing models often fail to adapt to new data domains without performance degradation.

Purpose of the Study:

• To develop a novel Domain-Incremental Weakly Supervised State-space Model (DIPathMamba) for pathology image segmentation.
• To enable segmentation using only image-level labels, overcoming the need for dense pixel-level annotations.
• To facilitate dynamic learning of new data domains while preserving performance on previously learned domains.

Main Methods:

• A shared feature extractor based on a hardware-aware state space model is designed.
• Multi-Instance Multi-Label Learning extracts pixel-level features, integrated into a Contrastive Mamba Block (CMB).
• A Domain Parameter Constraint Model (DPCM) and Collaborative Incremental Deep Supervision Loss (CIDSL) address incremental learning challenges and optimize parameter learning.

Main Results:

• The DIPathMamba model integrates fine-grained details with global context for improved segmentation.
• It generates more regionally consistent segmentation results compared to existing methods.
• Experimental results on three public datasets demonstrate superior performance over state-of-the-art approaches.

Conclusions:

• The proposed DIPathMamba effectively performs weakly supervised pathology image segmentation.
• It addresses the limitations of dense annotations and domain-specific data challenges in digital pathology.
• The method shows significant potential for advancing automated analysis in digital pathology workflows.