DPFR: Semi-supervised gland segmentation via density perturbation and feature recalibration

In recent years, semi-supervised methods have attracted considerable attention in gland segmentation of histopathological images, as they can substantially reduce the annotation data burden for pathologists. The most widely adopted approach is the Mean-Teacher framework based on consistency regularization, which exploits unlabeled data information through consistency regularization constraints. However, due to the morphological complexity of glands in histopathological images, existing methods still suffer from confusion between glands and background, as well as gland adhesion. To address these challenges, we propose a semi-supervised gland segmentation method based on Density Perturbation and Feature Recalibration (DPFR). Specifically, we first design a normalized flow-based density estimator to effectively model the feature density distributions of glands, contours, and background. The gradient information of the estimator is then exploited to determine the descent direction in low-density regions, along which perturbations are applied to enhance feature discriminability. Furthermore, a contrastive-learning-based feature recalibration module is designed to alleviate inter-class distribution confusion, thereby improving gland-background separability and mitigating gland adhesion. Extensive experiments on three public gland segmentation datasets demonstrate that the proposed method consistently outperforms existing semi-supervised approaches, achieving state-of-the-art performance with a substantial margin. The code repository address is https://github.com/Methow0/DPFR.