CTNet: A CNN-Transformer Dual-Branch Network for 2D-3D Rigid Registration

Moyamoya Disease is a complex cerebrovascular disorder, and surgeons diagnose it by establishing anatomical correspondence between 3D Time of Flight Magnetic Resonance Angiography (TOF-MRA) and 2D Digital Subtraction Angiography (DSA) cerebrovascular images. Existing 2D-3D registration algorithms in the field of medical imaging predominantly rely on convolutional neural networks (CNNs) and transformers as their core frameworks, and leverage feature learning to capture key information for precise alignment. However, the recent method combines CNNs and transformers in a sequential manner for global feature extraction, which results in insufficient global context modeling capability. Furthermore, it does not construct projections specifically for 3D TOF images, which leads to the loss of projection information and severely limits its registration performance. To alleviate these issues, we come up with a CNNtransformer dual-branch network (CTNet) for cross-modal 2D3D rigid registration, which evaluates the similarity between the global contextual information of the projected and fixed images, along with the correspondence of their local features, and utilizes the consistency between these two as a guiding criterion for optimizing spatial pose parameters during registration. In addition, we propose an improved projective spatial transformer (iProST) that employs parallel ray clusters to project instead of cone-shaped ray clusters, which preserves the critical information within TOF images. The CTNet and iProST are combined to form the registration method, and extensive experiments with an internal dataset demonstrate its effectiveness.