From Local to Global: An Anatomy-Aware Deep Learning Approach for Deformable Registration of Brain MR Images

Deformable registration serves as the foundation for longitudinal and population-based medical image analysis. In cross-sectional studies of brain magnetic resonance imaging (MRI), accurately aligning the regions of interest (ROIs) of various subjects, especially in the context of specific disease analysis, poses a challenge due to the complexity of brain structures and individual variations. The existing research predominantly utilizes the convolutional neural networks (CNNs) for learning global deformable registration, with little attention paid to the independent performance of neural networks on specific anatomical tissues. To address above problem, this paper proposed an anatomy-aware deep learning approach from local optimization to global information for medical deformable registration. In the method, the information of special brain anatomical tissue structure is selected to optimize the registration accuracy for the target region, and the anatomical perception constraint is designed to capture the information of these structures, so as to ensure the physiological rationality of the final registration deformation field. Compared with the recently developed state-of-the-art approaches, the proposed registration method can achieve the highest registration accuracy on both specific anatomical tissues and the overall image texture, and maintain image smoothness.