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

Colorectal cancer histopathological grading relies on the accurate segmentation of glandular structures. Current deep learning-based methods depend heavily on large-scale pixel-level annotations that are labor-intensive and not amenable to clinical practice. Weakly supervised semantic segmentation offers a promising alternative; yet, existing class activation map-based weakly supervised semantic segmentation approaches often produce incomplete, low-quality pseudo-masks that overemphasize discriminative regions and fail to provide reliable supervision for unannotated glandular structures, limiting their suitability for dense histopathology segmentation under sparse supervision. We propose a novel weakly supervised teacher-student framework that leverages sparse pathologists' annotations and an Exponential Moving Average-stabilized teacher network to generate refined pseudo-masks.

Our framework integrates confidence-based filtering, adaptive fusion of teacher predictions with limited ground truth, and curriculum-guided refinement, enabling the student network to progressively delineate and accurately segment unannotated glandular regions. We validated our framework on an institutional colorectal cancer cohort from The Ohio State University Wexner Medical Center, consisting of 60 hematoxylin and eosin-stained whole-slide images from independent patients with varying degrees of gland differentiation, as well as on public benchmarks including the Gland Segmentation dataset (derived from stage T3-T4 colorectal adenocarcinomas), TCGA-COAD, TCGA-READ, and SPIDER.

The proposed framework achieved strong performance on the institutional dataset despite limited annotations. On the Gland Segmentation dataset, it demonstrated competitive performance compared to both weakly and fully supervised approaches, achieving a mean Intersection over Union of 80.10% ± 1.52 and a mean Dice coefficient of 89.10% ± 2.10. Moreover, cross-cohort evaluations showed robust generalization on TCGA-COAD and TCGA-READ without requiring additional annotations, while reduced performance on SPIDER reflected pronounced domain shift.

Our framework provides an annotation-efficient and generalizable paradigm for accurate gland segmentation in colorectal histopathology, offering a practical pathway toward significantly reducing annotation burdens while preserving high segmentation fidelity.