Multi-resolution Coarse-to-fine Registration Approach for Liver Computed Tomography Image Analysis
Xuan Loc Pham, Quoc Anh Le, Duc Trinh Chu, Ha Manh Luu · 2022
Computed Tomography (CT) image contains vital medical information of patients and thus being irreplaceable in liver cancer treatment. Recently, computer-aided methods are increasingly applied into CT image processing, especially medical image segmentation and registration, and achieved promising results. However, performing non-rigid registration on liver CT images is challenging due to the large deformation caused by the big size of the liver organ. In this study, we propose a method for solving the liver registration problem, which utilizes convolutional neural network (CNN) with multi-resolution coarse-to-fine registration strategy to step-by-step deform the moving image to get closer the fixed shape. The proposed network is trained unsupervisedly for ease of expandability. We extensively evaluated the trained model on a variety of public liver datasets using dice (DSC), intersection over union (IoU) and landmark distance metrics, and compare to the performance of two well-known CNN-based registration methods. Experimental results show that the proposed method achieves promising results and proves its potential in the registration of CT images of diverse liver shapes.