History-Aware Multi-Expert-View Style Perturbation for Domain Adaptation in Clinical Medical Imaging Diagnosis

Mental health disorders, such as autism and anxiety, significantly affect individual well-being and public health. Automated segmentation of medical images, such as magnetic resonance imaging (MRI), is promising in the early diagnosis and prevention of these conditions. However, existing image analysis methods face inherent domain shift challenges in clinical applications due to variations in imaging centers, scanners, and patient populations. To address these limitations, this study proposes a history-aware multi-expert-view style perturbation model for domain adaptation in medical image segmentation. First, a multi-expert-view style perturbation module is proposed to dynamically generate diverse augmented features, enhancing the generalization ability and the capacity to learn domain-invariant characteristics. Second, a history-aware semantic alignment strategy is developed to construct diverse historical representations for each class, improving semantic consistency within classes and enhancing inter-class discriminability. Finally, a mutual AdaIN module is proposed to align style representations between the source and target domains, thereby reducing domain-specific style discrepancies. Extensive experiments conducted on two public datasets, including a brain dataset and a cardiac dataset, demonstrate superior performance of the proposed method in the field of unsupervised domain adaptation for medical image segmentation, underscoring its potential to enhance diagnostic accuracy and facilitate robust deployment in diverse clinical environments.