CONTRAST UNCERTAINTY DOMAIN ALIGNMENT FOR CROSS-DOMAIN PANCREATIC IMAGE SEGMENTATION

Multiple phase/modality images can provide more morphological and functional information about the pancreas for diagnosing pancreatic cancer. Cross-domain pancreatic image segmentation meets the demand for time-consuming manual annotation for multiple phase/modality images. However, the large domain discrepancy, individual difference and the large deformation make traditional methods lead to the instability of style transfer and shape deformation during domain transfer. To address the above issues, a novel domain adaptation network is proposed to improve the segmentation of the pancreas in the target phase/modality image. To ensure the stability of style transfer, features of the transformed images and target images are aligned by using an Attentional Feature Fusion Module (AFFM) based adversarial learning in feature space. To maintain the shape invariance, the uncertainty-constrained consistency loss is presented to constrain training of the proposed framework. The proposed framework is evaluated with two abdominal image datasets, and the experimental results show that it outperforms the state-of-the-art approaches.