Unsupervised cycle-consistent network for 3D pelvic mono- or multimodal deformable image registration

用于三维骨盆单模态或多模态可变形图像配准的无监督循环一致性网络

阅读:1

Abstract

BACKGROUND: To develop a novel deformable image registration method aimed at improving the accuracy of multimodal image registration in radiotherapy. METHODS: An unsupervised cycle-consistent network (CycleVOS) was proposed as the deformable image registration model, which integrates a cycle-consistent registration framework, a voxel optimization smooth (VOS) module, and a modality-independent neighbourhood descriptor (MIND) loss function. Two convolutional neural networks (CNNs) were employed to generate spatial deformable grids, ensuring consistency in the deformed images. In this study, the CycleVOS model was evaluated on multiple registration tasks, i.e., CT-CT registration, CT-CBCT registration and CT-MR registration tasks. The datasets consisted of 128 paired CT-MR images, 90 paired CT-CT images and 190 paired CT-CBCT images. Among these, 28 CT-MR, 25 CT-CT, and 40 CT-CBCT pairs were used as test data. The performance of the CycleVOS model was compared with that of the Elastix B-spline method using normalized mutual information (NMI) and the dice similarity coefficient (DSC) as evaluation metrics. RESULTS: In terms of the test data, the NMI values obtained with the CycleVOS model and the B-spline method were 0.211 vs. 0.205 for CT-MR images, 0.508 vs. 0.485 for CT-CT images, and 0.404 vs. 0.379 for CT-CBCT images. The DSCs for the CycleVOS model were consistently greater than those for the B-spline method across all propagated contours in CT-CT data (bladder, bone marrow, body, CTV, and PTV) and CT-CBCT data (bladder, spinal cord, left femoral head, right femoral head, and bone marrow). Furthermore, compared with the B-spline method, the CycleVOS model trained on mixed data achieved comparable or superior accuracy. CONCLUSIONS: The proposed CycleVOS model is highly accurate for both mono- and multimodal image registration. Its performance when trained on mixed data suggests that integrating various registration tasks into a single model is a feasible direction for future development. CLINICAL TRIAL NUMBER: Not applicable.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。