A semi-supervised computer-aided diagnosis network for reducing clinical annotation burden under limited annotation

Abstract

Objective. Recent advances in automatic medical image segmentation have achieved remarkable success; however, most existing approaches still heavily rely on large-scale pixel-level annotations provided by clinicians, making it difficult to effectively alleviate the clinical labeling burden. Semi-supervised medical image segmentation has therefore emerged as a promising solution for exploiting abundant unlabeled data under limited annotation settings. Nevertheless, despite significant progress in semi-supervised learning (SSL), the inherent discrepancy between pseudo-labels and ground-truth annotations leads to unreliable supervision, preventing SSL methods from fully replacing supervised learning. Moreover, existing approaches still struggle to address boundary ambiguity in medical images, which often causes inaccurate region delineation and erroneous predictions. Approach. To overcome these limitations, we propose FFT-EMatch, a novel self-knowledge distillation framework based on a multi-branch differential perturbation strategy. Specifically, a boundary-consistency regularization mechanism is introduced to explicitly decompose medical images into high- and low-frequency components for frequency-domain modeling, enabling dedicated consistency learning of structural boundary representations across heterogeneous perturbation views. Furthermore, an uncertainty-aware, entropy-driven pseudo-label recalibration and alignment strategy is developed to dynamically correct unreliable pseudo-label regions and provide supervision signals that are closer to ground-truth annotations during semi-supervised training. Main results. Extensive experiments on three public benchmark datasets demonstrate that FFT-EMatch consistently outperforms state-of-the-art semi-supervised medical image segmentation methods. Notably, on the BUSI and MS-CMRSeg 2019 datasets, the proposed framework achieves performance comparable to fully supervised training using only 25% and 20% labeled data, respectively. Significance. These results indicate that FFT-EMatch can effectively improve pseudo-label reliability and boundary-aware representation learning under limited annotation settings, highlighting its promise for annotation-efficient medical image segmentation and its potential applicability in practical medical image analysis scenarios.