Using Sparse Patch Annotation for Tumor Segmentation in Histopathological Images.

Area of Science:

• Digital pathology
• Medical image analysis
• Computational biology

Background:

• Accurate tumor segmentation in histopathology is crucial but demands extensive pixel-wise annotations.
• Fully-supervised methods require significant manual effort for annotation, hindering large-scale analysis.

Purpose of the Study:

• To develop a novel weakly supervised segmentation framework to reduce the annotation burden in histopathological image analysis.
• To leverage sparse patch annotations and semi-supervised learning for efficient tumor segmentation.

Main Methods:

• Proposed a weakly supervised segmentation framework utilizing sparse patch annotations (tumor/normal labels on small image portions).
• Introduced PSeger, a patch-wise segmentation model with dual branches for patch and image classification to enhance feature learning and reduce overfitting.
• Implemented an innovative semi-supervised strategy combining consistency learning and self-training to utilize unlabeled images.

Main Results:

• The framework achieved competitive performance compared to fully supervised pixel-wise segmentation models.
• Demonstrated effectiveness when trained on the BCSS dataset with only 25% of images sparsely labeled.
• The two-branch architecture of PSeger facilitated learning generalizable features from limited annotated data.

Conclusions:

• The proposed weakly supervised framework significantly reduces the manual annotation effort required for histopathological image segmentation.
• This approach shows great potential for improving the efficiency and scalability of tumor segmentation in digital pathology.
• Sparse patch annotation combined with semi-supervised learning offers a viable alternative to fully supervised methods.