Unsupervised Domain Adaptation in Medical Image Segmentation via Fourier Feature Decoupling and Multi-teacher Distillation

Unsupervised domain adaptation (UDA) has recently garnered widespread attention in the field of medical image segmentation by transferring knowledge from labeled source datasets to enhance model segmentation performance on unlabeled target domain data. Among these, reducing inter-domain discrepancies by aligning features between the source and target domains is a mainstream method in this field. These approaches typically consist of two stages: data alignment and training segmentation. However, they all face challenges: (1) the generated images during the data alignment stage are of poor quality, impacting the performance of the segmentation model; (2) during the training segmentation stage, the network is trained only with labeled synthetic target domain data, failing to fully utilize the potential of other generated data, thus limiting the improvement in model segmentation performance. To address these issues, we design a medical image unsupervised domain adaptation segmentation model, UDA-FMTD, based on Fourier feature decoupling and multi-teacher distillation. Evaluations conducted on the MICCAI 2017 MM-WHS cardiac dataset have demonstrated the effectiveness and superiority of this method.