A Comparative Study on Semi-Supervised Medical Image Segmentation Methods

Medical image segmentation is critical for disease detection and therapy, significantly contributing to the enhancement of diagnostic precision and therapeutic efficacy. Given the intrinsic complexity of medical images and the relative scarcity of annotated data, semi-supervised segmentation algorithms demonstrate unique advantages, providing effective solutions for medical image analysis. The purpose of this study is to evaluate the performance of four semi-supervised learning network models in medical image segmentation, including dual-task consistency network (DTC), mutual consistency network (MC-Net), smoothness and class separation network (SS-Net), and bidirectional copy-paste network (BCP), especially focusing on their performance in the case of limited annotated samples. To quantitatively analyze the performance of each model, evaluation metrics such as Dice coefficient, Jaccard coefficient, 95% Hausdorff distance (95HD), and Average Surface Distance (ASD) were employed. The dataset used in the experiment was sourced from the public dataset of the Atrial Segmentation Challenge. The experimental results indicated that MC-Net performed the best across all evaluation metrics, particularly excelling in Dice coefficient and 95HD, demonstrating significant advantages in segmentation accuracy and boundary delineation. BCP exhibited excellent performance in Jaccard coefficient and ASD, while DTC ranked closely to MC-Net in terms of Dice coefficient but slightly lagged in other metrics. SS-Net showed limitations in segmentation accuracy. Confronted with the scarcity of annotated data, this study establishes a theoretical groundwork for improving methodologies in medical image segmentation.