Semi-supervised Segmentation of Histopathology Images with Noise-Aware Topological Consistency

Area of Science:

• Digital Pathology
• Medical Image Analysis
• Computer Vision

Background:

• Accurate segmentation of densely packed structures like glands and nuclei is vital for digital pathology analysis.
• Manual pixel-wise annotation is labor-intensive, necessitating semi-supervised approaches for segmentation tasks.
• Current semi-supervised methods often produce topological inaccuracies, such as merged or missing structures.

Purpose of the Study:

• To develop the first semi-supervised segmentation method, TopoSemiSeg, capable of learning topological representations from unlabeled histopathology images.
• To address the challenge of noisy topological predictions from unlabeled data in semi-supervised learning.
• To improve the robustness and accuracy of segmentation models against topological errors.

Main Methods:

• Proposed TopoSemiSeg, a novel semi-supervised method for histopathology image segmentation.
• Introduced a noise-aware topological consistency loss to align teacher and student model representations.
• Decomposed prediction topology into signal and noisy components to learn true topological signals and enhance noise robustness.

Main Results:

• Demonstrated superior performance of TopoSemiSeg on public histopathology image datasets compared to existing methods.
• Achieved significant improvements, particularly on topology-aware evaluation metrics.
• Validated the method's effectiveness in learning accurate topological representations from unlabeled data.

Conclusions:

• TopoSemiSeg represents a significant advancement in semi-supervised segmentation for digital pathology.
• The proposed noise-aware topological consistency loss effectively mitigates topological errors in unlabeled data.
• The method shows strong potential for improving downstream analyses in digital pathology by providing accurate, topology-aware segmentations.