29 Computer Vision News Learn2Reg requires establishing a 2D or 3D displacement vector for each point in the fixed image, equating to two or three continuous labels per pixel or voxel. This vector bridges the fixed and moving images, meaning the solution exists not within either image, but in the space between them. Unlike for segmentation or classification, medical experts cannot annotate the ‘ground-truth’ for image registration. Despite this challenge, we have compiled a diverse dataset encompassing inter- and intra-patient, and single and multi-modal registration tasks. A substantial part of our training dataset includes annotations like segmentations or keypoints for supervised training. We also provide unannotated data for unsupervised or weakly-supervised training using cost functions from conventional registration methods. To evaluate the performance of registration methods, we use various auxiliary metrics. These metrics measure similarity through segmentation and keypoints, deformation plausibility, robustness, and runtime. We ensure that methods are only ranked higher if their results show a statistically significant improvement, accounting for random noise effects.
RkJQdWJsaXNoZXIy NTc3NzU=