Cyclic deformable medical image registration with prompt: deep fusion of diffeomorphic and transformer methods

Abstract

Medical image registration is a fundamental task in medical image analysis. Recently, competitive methods, such as deep learning-based registration and deformable registration, which have demonstrated promising results, have been proposed. However, meeting the demands for high precision in clinical applications is still a challenge. Here, we propose a cyclic optimization registration framework that deeply fuses diffeomorphic and deep learning methods through a single forward-two path structure. A neural network estimates the initial deformation field, which directly generates registered images to enhance feature extraction capabilities. Additionally, dynamic diffeomorphism is introduced for the initial deformation field to generate the final deformation field, ensuring the invertibility of the transformation. We incorporate the Dense Spatial Correspondence Prompt module for cyclically learning the final deformation field, enabling the network to estimate smoother and more accurate spatial transformations. Experiments conducted on a publicly available 3D dataset demonstrate exceptional registration accuracy with a DSC of 0.621 and an SSIM of 0.913, while preserving desirable diffeomorphic properties with almost zero non-positive Jacobians.