(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: registration is an important step in radiation treatment planning. Iterative calculation is the most common method in medical image registration, but it is relatively time-consuming. The goal of this study is to propose a model of using the Cycle-Consistent Fully Convolutional Network (FCN) for fast 3D CT-MR deformable registration.
Methods: deformation network firstly receives multimodal image pairs and outputs the deformed transformation. Then the MR image are deformed to get the deformed MR image. The CT image and deformed MR are subsequently input into the deformation network again to obtain the reconstructed transformation and reconstructed image pairs. We standardize all training image data to ensure the consistency of the distribution range of pixel values of all images. All the image pixel values are mapped to the range of (-1, 1). In terms of loss functions, we use regularization loss, cycle loss in CycleGAN and a metric called modality independent neighborhood descriptor (MIND) to perform deformable registration on CT-MR images.
Results: pelvic cases with MR and CT images are studied. Among them, sixty-four cases are used as training data set and ten cases are used as test data set. The performance of the proposed method is compared with that of Elastix software, MIM software and FCN. The results show that the proposed method achieved the best performance among the four registration methods in terms of registration accuracy and the method was more stable than others in general. In terms of average registration time, Elastix takes 64 seconds, MIM software takes 28 seconds and the proposed method is found to be significantly faster, taking less than 0.1 seconds.
Conclusion: proposed method not only ensures the accuracy of deformable image registration, but also greatly reduces the time required for image registration process.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
