A Two-Stage Learnable Framework for Liver MR-CT Registration under the Constraints of Arterial and Venous

Multimodal registration of liver MR and CT images is crucial for image-guided surgical planning of small hepatocellular carcinoma. However, giving all attention to the registration of organ volume contours might lead to a tolerance for inaccurate matching of important internal liver landmarks like arteries and veins, which are essential for pinpointing tumor positions during ablation. In order to register the liver and blood vessels simultaneously, we propose a two-stage unsupervised framework that estimates the deformation field for registering liver MR-CT images. The first phase involves affine registration of the image pair, followed by a coarse to fine deformable image registration to further refine the alignment. To enhance the registration process's stability, we employ structural features from image pairs as inputs for the network. Meanwhile, we develop a technique to identify matching points in liver veins in sets of images and create constraints based on liver vein information. This technique is successful in reducing the issue of abnormal deformation of the hepatic vein when performing volumetric registration. Quantitative and qualitative evaluations on liver MR-CT datasets show the priority of our method.