FMB: Dual-view fusion and registration of 2D DSA images and 3D MRA images for neurointerventional-based procedures

Objective: Alignment between preoperative images (high-resolution magnetic resonance imaging, magnetic resonance angiography) and intraoperative medical images (digital subtraction angiography) is currently required in neurointerventional surgery. Treating a lesion is usually guided by a 2D DSA silhouette image. DSA silhouette images increase procedure time and radiation exposure time due to the lack of anatomical information, but information from MRA images can be utilized to compensate for this in order to improve procedure efficiency. In this paper, we abstract this into the problem of relative pose and correspondence between a 3D point and its 2D projection. Multimodal images have a large amount of noise and anomalies that are difficult to resolve using conventional methods. According to our research, there are fewer multimodal fusion methods to perform the full procedure. Approach: Therefore, the paper introduces a registration pipeline for multimodal images with fused dual views is presented. Deep learning methods are introduced to accomplish feature extraction of multimodal images to automate the process. Besides, the paper proposes a registration method based on the Factor of Maximum Bounds (FMB). The key insights are to relax the constraints on the lower bound, enhance the constraints on the upper bounds, and mine more local consensus information in the point set using a second perspective to generate accurate pose estimation. Main results: Compared to existing 2D/3D point set registration methods, this method utilizes a different problem formulation, searches the rotation and translation space more efficiently, and improves registration speed. Significance: Experiments with synthesized and real data show that the proposed method was achieved in accuracy, robustness, and time efficiency.