Abstract
Teacher-student frameworks have emerged as a leading approach insemi-supervised medical image segmentation, demonstrating strong performanceacross various tasks. However, the learning effects are still limited by thestrong correlation and unreliable knowledge transfer process between teacherand student networks. To overcome this limitation, we introduce a novelswitching Dual-Student architecture that strategically selects the mostreliable student at each iteration to enhance dual-student collaboration andprevent error reinforcement. We also introduce a strategy of Loss-AwareExponential Moving Average to dynamically ensure that the teacher absorbsmeaningful information from students, improving the quality of pseudo-labels.Our plug-and-play framework is extensively evaluated on 3D medical imagesegmentation datasets, where it outperforms state-of-the-art semi-supervisedmethods, demonstrating its effectiveness in improving segmentation accuracyunder limited supervision.