ReUINet: a Fast GNL Distortion Correction Approach on a 1.0T MRI-Linac Scanner.

PURPOSE The hybrid system combining a magnetic resonance imaging (MRI) scanner with a linear accelerator (Linac) has become increasingly desirable for tumour treatment because of excellent soft-tissue contrast and non-ionising radiation. However, image distortions caused by gradient non-linearity (GNL) can have detrimental impacts on real-time radiotherapy using MRI-Linac systems, where accurate geometric information of tumours is essential. METHODS In this work, we proposed a deep convolutional neural network-based method to efficiently recover undistorted images (ReUINet) for real-time image guidance. The ReUINet, based on the encoder-decoder structure, was created to learn the relationship between the undistorted images and distorted images. The ReUINet was pre-trained and tested on a publically available brain MR image dataset acquired from 23 volunteers. Then, transfer-learning was adopted to implement the pre-trained model (i.e., network with optimal weights) on the experimental 3D grid phantom and in-vivo pelvis image datasets acquired from the 1.0 T Australian MRI-Linac system. RESULTS Evaluations on the phantom (768 slices) and pelvis data (88 slices) showed that the ReUINet achieved over 15-times and 45-times improvement on computational efficiency in comparison with standard interpolation and GNL-encoding methods, respectively. Moreover, qualitative and quantitative results demonstrated that the ReUINet provided better correction results than the standard interpolation method, and comparable performance compared to the GNL-encoding approach. CONCLUSIONS Validated by simulation and experimental results, the proposed ReUINet showed promise in obtaining accurate MR images for the implementation of real-time MRI-guided radiotherapy.