Semi-Supervised Instance Segmentation in Whole Slide Images via Dense Spatial Variability Enhancing.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Current whole slide image (WSI) segmentation focuses on tumor vs. background.
• Segmenting distinct tumor instances in WSIs with limited annotations is challenging and underexplored.

Purpose of the Study:

• To formally propose and evaluate semisupervised instance segmentation (Semi-IS) for WSIs.
• To address the challenge of learning from unlabeled data for improved instance segmentation.

Main Methods:

• Developed a semisupervised instance segmentation (Semi-IS) framework using contrastive learning.
• Treated image patches as composed tokens for learning intra-class similarity and inter-class dissimilarity.
• Incorporated noise elimination and preservation techniques to refine segmented instance boundaries.

Main Results:

• Semi-IS achieved near fully-supervised state-of-the-art performance using only 30% annotated data in clinical multi-instance segmentation tasks.
• Demonstrated improved segmentation accuracy by approximately 2% on public cell pathology datasets.
• Validated the effectiveness and generalizability of the Semi-IS approach across histopathology and cellular pathology.

Conclusions:

• Semisupervised instance segmentation (Semi-IS) is a viable and effective approach for WSIs.
• Semi-IS significantly reduces the need for extensive manual annotations in digital pathology.
• The proposed framework shows strong potential for clinical applications in pathology image analysis.