Semi-supervised medical image segmentation based on cross-teaching between CNN and Transformer
Due to the lack of high-quality annotations in the medical imaging field,semi-super-vised learning methods have gained significant attention for image semantic segmentation tasks.This paper proposes a method that leverages both Convolutional Neural Network(CNN)and Transformer through cross-teaching.The method simplifies the classical deep co-training by replacing consistency regularization with cross-teaching.It utilizes a cyclic pseu-do-labeling scheme,converting prediction differences into unsupervised losses.This encoura-ges consistent low-entropy predictions from both networks.The proposed method is experi-mentally validated on the ISIC 2018 dataset,achieving Dice coefficients of 87.25％and Jac-card coefficients of 79.17％with a 20％annotation ratio.Compared to the supervised U-Net++training results,there is an improvement of 2.89％and 3.53％,respectively.The meth-od surpasses mainstream semi-supervised approaches,demonstrating its effectiveness in med-ical image segmentation.
万方数据
维普
