Search

Purpose: This study aims to evaluate the impact of key parameters on the pseudo computed tomography (pCT) quality generated from magnetic resonance imaging (MRI) with a 3-dimensional (3D) convolutional neural network., Methods and Materials: Four hundred two brain tumor cases were retrieved, yielding associations between 182 computed tomography (CT) and T1-weighted MRI (T1) scans, 180 CT and contrast-enhanced T1-weighted MRI (T1-Gd) scans, and 40 CT, T1, and T1-Gd scans. A 3D CNN was used to map T1 or T1-Gd onto CT scans and evaluate the importance of different components. First, the training set size's influence on testing set accuracy was assessed. Moreover, we evaluated the MRI sequence impact, using T1-only and T1-Gd-only cohorts. We then investigated 4 MRI standardization approaches (histogram-based, zero-mean/unit-variance, white stripe, and no standardization) 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, 2 networks, namely HighResNet and 3D UNet, were compared to evaluate the architecture's impact on the pCT quality. The mean absolute error, gamma indices, and dose-volume histograms were used as evaluation metrics., Results: Generating models using all the available cases for training led to higher pCT quality. The T1 and T1-Gd models had a maximum difference in gamma index means of 0.07 percentage point. The mean absolute error obtained with white stripe was 78 ± 22 Hounsfield units, which slightly outperformed histogram-based, zero-mean/unit-variance, and no standardization (P < .0001). Regarding the network architectures, 3%/3 mm gamma indices of 99.83% ± 0.19% and 99.74% ± 0.24% were obtained for HighResNet and 3D UNet, respectively., Conclusions: Our best pCTs were generated using more than 200 samples in the training data set. Training with T1 only and T1-Gd only did not significantly affect performance. Regardless of the preprocessing applied, the dosimetry quality remained equivalent and relevant for potential use in clinical practice., (Copyright © 2020 Elsevier Inc. All rights reserved.)

Radiomics relies on the extraction of a wide variety of quantitative image-based features to provide decision support. Magnetic resonance imaging (MRI) contributes to the personalization of patient care but suffers from being highly dependent on acquisition and reconstruction parameters. Today, there are no guidelines regarding the optimal pre-processing of MR images in the context of radiomics, which is crucial for the generalization of published image-based signatures. This study aims to assess the impact of three different intensity normalization methods (Nyul, WhiteStripe, Z-Score) typically used in MRI together with two methods for intensity discretization (fixed bin size and fixed bin number). The impact of these methods was evaluated on first- and second-order radiomics features extracted from brain MRI, establishing a unified methodology for future radiomics studies. Two independent MRI datasets were used. The first one (DATASET1) included 20 institutional patients with WHO grade II and III gliomas who underwent post-contrast 3D axial T1-weighted (T1w-gd) and axial T2-weighted fluid attenuation inversion recovery (T2w-flair) sequences on two different MR devices (1.5 T and 3.0 T) with a 1-month delay. Jensen-Shannon divergence was used to compare pairs of intensity histograms before and after normalization. The stability of first-order and second-order features across the two acquisitions was analysed using the concordance correlation coefficient and the intra-class correlation coefficient. The second dataset (DATASET2) was extracted from the public TCIA database and included 108 patients with WHO grade II and III gliomas and 135 patients with WHO grade IV glioblastomas. The impact of normalization and discretization methods was evaluated based on a tumour grade classification task (balanced accuracy measurement) using five well-established machine learning algorithms. Intensity normalization highly improved the robustness of first-order features and the performances of subsequent classification models. For the T1w-gd sequence, the mean balanced accuracy for tumour grade classification was increased from 0.67 (95% CI 0.61-0.73) to 0.82 (95% CI 0.79-0.84, P = .006), 0.79 (95% CI 0.76-0.82, P = .021) and 0.82 (95% CI 0.80-0.85, P = .005), respectively, using the Nyul, WhiteStripe and Z-Score normalization methods compared to no normalization. The relative discretization makes unnecessary the use of intensity normalization for the second-order radiomics features. Even if the bin number for the discretization had a small impact on classification performances, a good compromise was obtained using the 32 bins considering both T1w-gd and T2w-flair sequences. No significant improvements in classification performances were observed using feature selection. A standardized pre-processing pipeline is proposed for the use of radiomics in MRI of brain tumours. For models based on first- and second-order features, we recommend normalizing images with the Z-Score method and adopting an absolute discretization approach. For second-order feature-based signatures, relative discretization can be used without prior normalization. In both cases, 32 bins for discretization are recommended. This study may pave the way for the multicentric development and validation of MR-based radiomics biomarkers.