Weakly Supervised Teacher-Student Framework with Progressive Pseudo-mask Refinement for Gland Segmentation.

Area of Science:

• Digital pathology
• Computational oncology
• Medical image analysis

Background:

• Accurate segmentation of glandular structures is crucial for colorectal cancer histopathological grading.
• Current deep learning methods require extensive pixel-level annotations, which are labor-intensive and impractical for clinical settings.
• Existing weakly supervised methods produce low-quality pseudo-masks, limiting their effectiveness for dense histopathology segmentation.

Purpose of the Study:

• To develop an annotation-efficient weakly supervised semantic segmentation framework for colorectal cancer glands.
• To generate refined pseudo-masks using sparse annotations and a stabilized teacher network.
• To improve the accuracy and reliability of gland segmentation in histopathology.

Main Methods:

• Proposed a novel weakly supervised teacher-student framework.
• Integrated confidence-based filtering, adaptive fusion of teacher predictions with limited ground truth, and curriculum-guided refinement.
• Validated the framework on an institutional colorectal cancer cohort and public benchmarks (Gland Segmentation dataset, TCGA-COAD, TCGA-READ, SPIDER).

Main Results:

• Achieved strong performance on the institutional dataset with limited annotations.
• Demonstrated competitive performance on the Gland Segmentation dataset, with a mean IoU of 80.10% and Dice coefficient of 89.10%.
• Showed robust generalization on TCGA-COAD and TCGA-READ datasets without additional annotations.

Conclusions:

• The proposed framework offers an annotation-efficient and generalizable solution for accurate gland segmentation in colorectal histopathology.
• It significantly reduces annotation burdens while maintaining high segmentation fidelity.
• Provides a practical pathway for clinical application in digital pathology.