Multi-scale Mean Teacher for Unsupervised Cross-Modality Abdominal Organ Segmentation with Limited Annotations

The achievement of Deep Convolutional Neural Networks (DCNNs) in abdominal organ segmentation can be attributed to the availability of extensive annotated data. Unsupervised Domain Adaptation (UDA) has been developed to achieve cross-modality image segmentation by utilizing well-annotated data from the source domain and large amounts of unlabeled data from the target domain. However, UDA still encounters performance degradation when there is a lack of annotations in the source domain. To address this issue, our study delves into a challenging UDA scenario where source domain annotations are limited. The proposed framework consists of two stages. Firstly, Contrastive Learning for Unpaired Image-to-Image Translation (CUT) and Cycle-Consistent Adversarial Networks (CycleGAN) are utilized to generate two sets of synthesized CT volumes from MRI data. These synthesized volumes, with distinct details and appearances, serve as training data to improve supervised segmentation. In the second stage, the generated images are fed into a self-ensemble learning network that bridges the domain gap using multi-scale supervised loss and uncertainty rectified consistency loss. By employing the Data Augmentation (DA) strategy, our proposed method has achieved superior performance in abdominal organs segmentation compared to other state-of-the-art segmentation methods, even with limited labeled source data. This improvement was observed across specific evaluation metrics, achieving an average Dice of 69.2 \(\%\) and an average ASD of 3.4 for cross-modality images.