The semi-supervised medical image segmentation method based on adaptive perturbation enhancement and progressive consistency propagation

Semi-supervised learning (SSL) has demonstrated strong potential in medical image segmentation by leveraging limited labeled data alongside abundant unlabeled data. However, its effectiveness is often constrained by unreliable pseudo-labels and inefficient use of unlabeled samples. To address these challenges, we propose a novel SSL framework to enhance segmentation accuracy and robustness under limited supervision. This framework integrates three components: a perception-guided adaptive perturbation module (PAPE), a consistency-aware region propagation module (CRP), and a progressive pseudo-label optimization strategy (PPO). The PAPE module, built upon a perception-guided modified simplified pulse-coupled neural network (PG-MSPCNN), generates multiscale perturbed images and adaptively adjusts the perturbation intensity based on sample complexity. The CRP module improves pseudo-label quality through adaptive pixel-wise consistency mapping and shape-aware region propagation. The PPO strategy follows a curriculum learning paradigm to gradually incorporate high-confidence samples, thereby enhancing model generalization. We evaluate our method on three benchmark datasets: Breast Ultrasound Images (BUSI), International Skin Imaging Collaboration (ISIC), and Diabetic Dermoscopic Telemedicine Images (DDTI), under both 10% and 20% labeled data settings. Our method consistently outperforms recent state-of-the-art methods across multiple metrics. For instance, on ISIC with 20% labeled data, our approach improves the Dice Index by 1.3% and the Jaccard Index by 1.89% compared to the best-performing baseline, demonstrating its effectiveness and strong generalization capability in low-supervision medical image segmentation. Our code is publicly available at: https://github.com/Nazn65/APC-Net.