Progressive and Coarse-to-Fine Network for Medical Image Registration Across Phases, Modalities and Patients

In this paper, we apply our proposed PCNet [12] on three different registration tasks assigned by the Learn2Reg challenge 2021 [1,5], i.e., CT-MR thorax-abdomen registration [3,1144], lung inspiration-expiration registration [i] and whole brain registration [1,13], well covering three key demands in clinical practice, i.e., registration across modalities, across phases, and across patients. In these tasks, an accurate and reasonable deformation field plays a crucial role while it is often difficult to estimate in large misalignments. The core conception of our PCNet is to decompose the target deformation field into multiple sub-fields in both progressive and coarse-to-fine manners, which dramatically simplifies the direct estimation of deformation field and thus leads to a robust registration performance. The evaluation results on the three tasks demonstrate a competitive performance of PCNet and its great scalability to meet various registration demands.