Dosimetry-driven quality measure of brain pseudo Computed Tomography generated from deep learning for MRI-only radiotherapy treatment planning.
Researchers
Journal
Modalities
Models
Abstract
This study aims at evaluating the impact of key parameters on the pseudo Computed Tomography (pCT) quality generated from Magnetic Resonance Imaging (MRI) with a 3D convolutional neural network (CNN).
402 brain tumor cases were retrieved yielding to associations of 182 Computed Tomography (CT)/T1 weighted MRI (T1), 180 CT/contrast enhanced T1 weighted MRI (T1-Gd) and 40 CT/T1/T1-Gd. A 3D CNN was used to map T1 or T1-Gd into CT and evaluate the importance of different components. First, the training set size influence on the testing set accuracy was assessed. Moreover, we evaluated the MR sequence impact, using T1 only and T1-Gd only cohorts. Then, we investigated four MRI standardization approaches, namely histogram-based (HB), zero-mean/unit-variance (ZMUV), White Stripe (WS) and no standardization (NS) based on training, validation and testing cohorts composed of 242, 81 and 79 patients cases respectively, as well as a bias field correction influence. Finally, two networks, namely HighResNet and 3D UNet, were compared to evaluate the architecture impact on the pCT quality. The Mean Absolute Error (MAE), gamma indices and dose volume histograms were used as evaluation metrics.
Generating models using all the available cases for training led to higher pCT quality. The T1 and T1-Gd models indicated maximum differences in gamma indices means of 0.07 percent point. The MAE obtained with WS was 78 Hounsfield Units (HU) +/-22HU, which slightly outperformed HB, ZMUV and NS (p<0.0001). Regarding the network architectures, 3%/3mm gamma indices of 99.83%+/-0.19% and 99.74%+/-0.24% were obtained for HighResNet and 3D UNet respectively.
Our best pCT were generated using more than 200 samples in the training dataset, while training with T1 only and T1-Gd only did not significantly affect the performance. Regardless of the preprocessing applied, the dosimetry quality remained equivalent and relevant for a potential use in clinical practice.
Copyright © 2020. Published by Elsevier Inc.