MRI-based Proton Radiotherapy for Prostate Cancer Using Deep Convolutional Neural Networks

Abstract

The purpose of this study is to develop a deep-learning-based method to generate synthetic CT (sCT) from anatomical MRI for potential MRI-only based prostate proton beam radiotherapy. We propose a 3D cycle-generative adversarial network (cycle-GAN) framework with dense blocks to effectively generate sCT from MRI. Cycle-GANs not only train the network from MRI to sCT, but also inversely reinforce the algorithm by generating cycle-MRI from sCT. To minimize the differences between the synthetic/cycle images and their respective real datasets, both transformations are implemented by a generator and the outputs are judged by a discriminator. Training the generator takes into account the estimation error loss as well as the discriminator feedback, which can make the generator’s output more similar to the real dataset and improve the discriminator’s ability to differentiate synthetic/cycle from real images. In the synthetic stage, the patches from testing MRI are fed into the well-trained transformation model to estimate the sCT patches. The full sCT is then generated through a patch fusion process. The sCT generation method was evaluated using 20 prostate cancer patients with 17 proton treatment plans. Our sCTs were registered to CT for generating MRI-based proton treatment plan. Mean absolute error (MAE) and normalized cross-correlation (NCC) were used to quantify the differences between the sCT and CT. Clinically-relevant dose volume histogram (DVH) metrics were extracted from CT- and sCT-based proton plans for quantitative dosimetric evaluation. Gamma analyses were performed for the comparison of planar dose distributions. Proton distal range displacement, and the individual pencil beam Bragg peak shift between both proton plans were also evaluated. The mean MAE and NCC between the sCT and CT were 50.89±16.79 HU and 0.93±0.03 for the entire cohort. Relative differences of the prostate DVH metrics between sCT and CT were less than 1%. Mean values of dose difference, absolute dose difference (in percent of the prescribed dose) were -0.07±0.07% and 0.23±0.08%. Mean gamma analysis pass rate was 99% using 3mm/3% criteria with 10% dose threshold. Median and mean absolute maximum range differences were 0.9 mm and 2.3 mm. Median and mean proton pencil beam spot Bragg peak shifts among all patients were 0.9 mm and 1.8 mm. We developed a novel deep-learning-based approach to generate sCT from routine MR for MRI-based proton radiotherapy. The proposed method reliably generated sCT from the MRI that provided dose calculation of comparable accuracy to the CT for proton treatment planning. Results of image quality, gamma analysis passing rate, DVH metrics, distal range displacement, and the individual pencil beam Bragg peak shift between two proton plans support further development of an MRI-only workflow for prostate proton radiotherapy.