面向医学图像分割的半监督自相似聚类方法OA

To address the issue of insufficient model generalization caused by the scarcity of labeled data and high labeling costs in medi-cal image segmentation,we propose a novel semi-supervised multi-class self-similarity clustering framework(SMSC).Firstly,to fully leverage the inherent anatomical structure priors in unlabeled data,global features are dynamically divided into medial and lateral anatomi-cal subregions,enhancing the model's ability to represent local fine-grained features.Secondly,a multi-class self-similarity clustering module is proposed,which generates pseudo-labels for unlabeled data through unsupervised clustering and introduces a triplet loss con-straint to achieve intra-class compactness and inter-class separability in the feature space,mitigating model degradation under limited su-pervision.Finally,a pseudo-label collaborative training mechanism is constructed based on a clustering-guided strategy.Through joint optimization of a small amount of real annotations and a large number of high-quality pseudo-labels,effective propagation of labeling in-formation and steady improvement in model performance are achieved.SMSC significantly improves segmentation accuracy on knee MRI datasets(OAI-ZIB,OAI-Imorphics)and lung CT datasets(LUNA16),with Dice coefficients showing substantial improvement over the baseline nnU-Net model in multiple key structures,providing an effective solution for medical image segmentation under limited labeling resources.