3D Deformable Convolution for Medical Image Registration

Tong Huang, Nan Mu, Haoyang Xing, Xiaoning Li

Abstract: Medical image registration is a fundamental task in computeraided medical diagnosis. In recent years, researchers have proposed many deep learning-based registration models. Such methods can capture local feature information in images, significantly improving registration speed. However, these registration models have insufficient learning capabilities when dealing with complex deformations in medical images and cannot utilize multi-scale contextual information, leading to suboptimal registration accuracy. In this paper, we present DF-Net, a registration model that utilizes 3D deformable convolution to tackle these challenges. First, we implement the 3D deformable convolution algorithm to replace the original standard convolution, enhancing the registration network’s capacity to detect complex deformation features in medical images. Then, we design a Ladder Feature Fusion Module (Ladder-FFM) to densely fuse semantic information with deformation features in the context, improving the model’s capacity to recover deformations in medical images. Finally, We assessed our method using two publicly accessible brain MRI scan datasets. Theoretical analysis and experiments show that our registration model effectively handles complex deformations in medical images. Compared to existing methods, there is a certain improvement in registration accuracy, demonstrating the effectiveness of 3D deformable convolution in medical image registration.