Image registration using MR-based synthetic CT (sCT) generated by cycle-consistent adversarial networks

Image registration involves aligning multiple images within a common coordinate system to determine their geometric transformations. This study aims to improve diagnostic accuracy and efficiency by applying deep learning-based image registration between CT and MR images. Initially, the iterative closest point (ICP) technique was utilized to extract point clouds from CT and MR images and their corresponding segmentation masks. Through ICP-based alignment, the Dice Similarity Coefficient (DSC) for the segmentation mask (specifically, the femur head) improved from 0.29 to 0.91, and the Root Mean Square Error (RMSE) also decreased. However, to achieve more precise registration, a Cycle-GAN-based generative model was employed to synthesize CT (sCT) images from MR images, enabling registration to be performed on modality-unified images. The generated sCT images demonstrated high similarity to actual CT images, as indicated by a PSNR of 20.57 and an NCC of 0.93. Subsequently, registered between the MR images and sCT images yielded to a PSNR of 12.95 and an NCC of 0.62, indicating strong alignment with the CT images. This study addresses the inherent challenges of multi-modality image registration and highlights the effectiveness of utilizing unified synthetic images for improved registration performance. Future research will focus on enhancing data diversity and quality, as well as refining deep learning model architectures to further advance registration accuracy. These advancements are expected to contribute to the development of clinically applicable tools, utilizing improving the precision of medical image analysis and diagnosis.