Label-driven magnetic resonance imaging (MRI)-transrectal ultrasound (TRUS) registration using weakly supervised learning for MRI-guided prostate radiotherapy
Qiulan Zeng,Yabo Fu,Zhen Tian,Yang Lei,Yupei Zhang,Tonghe Wang,Hui Mao,Tian Liu,Walter J Curran,Ashesh B Jani,Pretesh Patel,Xiaofeng Yang
DOI: https://doi.org/10.1088/1361-6560/ab8cd6
IF: 3.5
2020-06-26
Physics in Medicine and Biology
Abstract:Registration and fusion of magnetic resonance imaging (MRI) and transrectal ultrasound (TRUS) of prostate can provide guidance for prostate brachytherapy. However, accurate registration remains a challenging task due to the lack of ground-truth regarding voxel-level spatial correspondence, limited field of view, low contrast-to-noise ratio in TRUS. In this study, we proposed a weakly supervised deep learning approach to address these issues.We employed deep learning techniques to combine image segmentation, affine and non-rigid registration to perform a deformable MRI-TRUS registration. First, we trained two separate fully convolutional neural networks to perform MRI and TRUS prostate segmentation. Then, a convolutional neural network was used to rigidly register MRI-TRUS images via affine registration. Third, a UNET-like network was applied for non-rigid registration. For both affine and non-rigid registration. Due to the unavailability of ground truth correspondences and the lack of accurate intensity-based image similarity measures, we propose to use prostate label-derived volume overlaps and surface agreements as an optimization objective function for weakly supervised network training. We proposed a hybrid loss function that integrated Dice and surface-based loss, and a bending energy regularization for the non-rigid registration.36 sets of patient data were used to test our registration method. Image registration results showed that the deformed MR image aligned well with the TRUS image, as judged by corresponding cysts and calcifications in prostate. Our method produced a mean target registration error (TRE) of 2.53 ±1.39 mm and a mean Dice of 0.91±0.02. The mean surface distance (MSD) and Hausdorff distance (HD) between the registered MR prostate shape and TRUS prostate shape were 0.88 mm and 4.41 mm, respectively.This work presents a deep learning-based, weakly supervised network for accurate MRI-TURS image registration. Our proposed method has achieved promising registration performance in terms of Dice, TRE, MSD and HD.
engineering, biomedical,radiology, nuclear medicine & medical imaging