Semi-supervised medical image segmentation via Cross Reliable Knowledge Distillation.
TL;DR
This work introduces a Class Semantic Distillation module, which transfers class-level semantic knowledge from labeled data to unlabeled representations through class prototype alignment in the feature space, facilitating discriminative feature learning, and develops a Low-Entropy Consistency module that dynamically emphasizes reliable low-entropy regions.
Abstract
Semi-supervised medical image segmentation aims to alleviate the dependence on large-scale annotations by exploiting unlabeled data. However, existing methods often lack effective mechanisms for leveraging semantic knowledge from labeled data and enabling reliable knowledge interaction among unlabeled perturbation branches. To address this, we propose Cross Reliable Knowledge Distillation (CRKD), a semi-supervised framework for medical images. Specifically, we first introduce a Class Semantic Distillation (CSD) module, which transfers class-level semantic knowledge from labeled data to unlabeled representations through class prototype alignment in the feature space, facilitating discriminative feature learning. Then, we propose a Low-Entropy Consistency (LEC) module that dynamically emphasizes reliable low-entropy regions, improving pixel-level prediction consistency under diverse perturbations. Finally, we develop a Cross Knowledge Distillation (CKD) module, consisting of Cross Pixel Knowledge Distillation (CPKD) and Cross Feature Knowledge Distillation (CFKD), which enables reliability-aware bidirectional knowledge transfer in both feature and pixel spaces, allowing perturbation branches to collaboratively correct uncertain representations and predictions without introducing additional network architectures. Extensive experiments demonstrate the effectiveness of CRKD, achieving significant performance gains and robust generalization across diverse imaging modalities and lesion types. Code is available at https://github.com/eeaesa/CRKD.