Adaptive Multi-Scale Deep Feature Fusion Network for Joint Cardiac Registration and Segmentation

To achieve an effective representation of motion in cardiac anatomical structures, we propose an end-to-end adaptive multi-scale deep feature fusion network (AMF-Reg). AMF-Reg consists of two components: a registration branch and a segmentation branch. The registration branch employs an adaptive multi-scale feature fusion strategy, which enhances the model's ability to capture both local and global features by integrating multi-frequency and multi-scale information. The segmentation branch shares an encoder with the registration branch and utilizes the generated motion field information to perform anatomical structure segmentation. Simultaneously, weak supervision constraints are imposed on the registration branch to further optimize its performance, thereby improving the model's capacity for anatomical structure localization and topology preservation. Experimental results indicate that our method achieves state-of-the-art performance in terms of registration accuracy and model generalization across both ACDC and CAP datasets, thereby underscoring its potential for application in clinical medical imaging. Our code is available at https://github.com/Pipixia911/AMF-Reg.