Self-Supervised Structure-Preserved Image Registration Framework for Multimodal Retinal Images

Abstract

Multi-modal image registration of ophthalmology images is vital for disease diagnosis and treatment plans. However, it is challenging as the divergences of image appearance, resolution, and different transformations among different modal images. Therefore, we propose an image registration framework for multimodal retinal images, which directly solves both rigid and deformable transformation. Considering the blood vessel should be consistent among different modal images, we propose a Structure-preserved registration network (SPR-Net) in the framework. Specifically, SPR-Net adopts structure-preserved modal transformation to provide generated multimodal images for the training of the registration network. We also propose a smooth loss function for the constraint of the predicted deformation field. Extensive experiments prove the effectiveness of our proposed registration framework.