Distribution-Constrained Semi-Supervised Learning for Medical Organ Segmentation

Abstract

Medical organ segmentation plays a crucial role in disease diagnosis and treatment planning. However, most existing methods focus primarily on the segmentation objects themselves while overlooking the positional correlations within images. These correlations can effectively align object features and provide boundary constraints for segmentation tasks. To address this limitation, this paper proposes a Distribution-Constrained Semi-Supervised Segmentation Model (DCSSM). Firstly, a Category Center of Mass Calculation Module (CCMCM) is introduced to capture positional correlations in medical images. Secondly, the model employs a shared encoder and a distribution network, which are pretrained using unlabeled samples to align input image features. The aligned feature boundaries then serve as constraints to guide mask generation during segmentation. Additionally, a Lightweight Multiscale Feature Extraction Module (LMFEM) is designed to efficiently transmit multiscale features to the decoder. Experimental results demonstrate that DCSSM achieves the best performance across multiple segmentation models on 5 medical segmentation datasets.