Towards Point Cloud-Based Medical Image Registration for Dynamic 4D-CT Imaging

Abstract

Dynamic computerized tomography (4D-CT) enables detailed analysis of musculoskeletal (MSK) joint motion. Estimating useful kinematic information from these images requires a registration of the obtained images. This study proposes a point-based registration method utilizing pre-segmented 4D-CT knee joint images to generate point clouds, employing 3D local deep descriptors (DIPs) encoded via a pre-trained PointNet-based deep neural network. We ask whether a network trained on indoor and outdoor datasets can effectively generalize to anatomical structures. The method was compared to traditional intensity-based registration using Target Registration Error (TRE) as a metric. Our evaluation involved registrations from different subjects, focusing on various anatomical structures of the knee, including the femur, tibia, and patella. The mean TRE for the intensity-based method was 2.29 ± 0.80 mm, while the point-based method achieved a mean TRE of 2.26 ± 0.73 mm. These results indicate that the point-based method offers comparable accuracy to intensity-based methods while reducing computational time. Although both methods require sequential registration across all timestamps, the point-based approach avoids the failures with distant timestamps encountered by the intensity-based method, which requires proper initialization. Additionally, the proposed method provides reliable extraction of kinematic parameters which have potential in understanding joint motion and MSK disorders.