Affiliations: [a] School of Mathematical Sciences, Zhejiang University, Hangzhou, China
| [b] College of Mathematics and Computer Science, Zhejiang A&F University, Hangzhou, China
| [c] China Mobile (Hangzhou) Information Technology Co., Ltd. | [d] Zhejiang Qiushi Institute for Mathematical Medicine, Hangzhou, China
Abstract: BACKGROUND:Deformable image registration (DIR) plays an important part in many clinical tasks, and deep learning has made significant progress in DIR over the past few years. OBJECTIVE:To propose a fast multiscale unsupervised deformable image registration (referred to as FMIRNet) method for monomodal image registration. METHODS:We designed a multiscale fusion module to estimate the large displacement field by combining and refining the deformation fields of three scales. The spatial attention mechanism was employed in our fusion module to weight the displacement field pixel by pixel. Except mean square error (MSE), we additionally added structural similarity (ssim) measure during the training phase to enhance the structural consistency between the deformed images and the fixed images. RESULTS:Our registration method was evaluated on EchoNet, CHAOS and SLIVER, and had indeed performance improvement in terms of SSIM, NCC and NMI scores. Furthermore, we integrated the FMIRNet into the segmentation network (FCN, UNet) to boost the segmentation task on a dataset with few manual annotations in our joint leaning frameworks. The experimental results indicated that the joint segmentation methods had performance improvement in terms of Dice, HD and ASSD scores. CONCLUSIONS:Our proposed FMIRNet is effective for large deformation estimation, and its registration capability is generalizable and robust in joint registration and segmentation frameworks to generate reliable labels for training segmentation tasks.
