Deformable adversarial registration network with multiple loss constraints

Abstract Deformable medical image registration has the necessary value of theoretical research and clinical application. Traditional methods cannot meet clinical application standards in terms of registration accuracy and efficiency. This article proposes a deformable generate adversarial registration framework, which avoids the dependence on ground-truth deformation. The proposed residual registration network based on Nested U-Net has excellent feature extraction ability and robustness. Multiple constraints that incorporate the potential information of anatomical segmentation extracted by the discriminator can help the model adapt to different modal registration tasks. Through interpatient X-ray chest registration, the deep-supervised training method, and the proposed loss constraint are proved to improve the model's performance and training stability. The experimental results show that our model, compared with state-of-the-art methods, provides a more accurate spatial alignment relationship between different patients’ lung organs while ensuring the displacement field's authenticity. Finally, we explored the relationship between the accuracy and validity of the model.