Adaptive segmentation of liver image domain by combining image alignment and feature alignment

Liver image segmentation is a critical task in medical image processing. However, significant performance degradation often occurs due to distribution differences between different imaging modalities. To address the domain adaptation challenge in cross-modal liver segmentation, this paper proposes an unsupervised domain adaptation method that combines image alignment and feature alignment. Specifically, we design a unified end-to-end training framework. First, at the input level, an unpaired image translation module is employed to reduce the distribution gap between domains. Then, an adversarial training strategy is introduced to align feature distributions between the source and target domains, thereby enhancing the network's generalization capability. To further improve the alignment of low-level features, a supervision mechanism is incorporated to directly guide the model training, strengthening the alignment effect. Additionally, to increase the focus on the target region, we integrate the Attention U-Net architecture into the segmentation network and optimize the segmentation performance using a hybrid loss function that combines cross-entropy loss and Dice loss.