Deep Neural Networks for Radiographic Risk Stratification of Gliomas

Abstract

Deep learning algorithms allow for non-invasive risk stratification of gliomas based on radiographic data, which may help guide clinical decision making and further our understanding of glioma subtypes. The aim of this study was to accurately classify histopathologic grade based on quantitative imaging features on MRI through the development a convolutional neural network, as well as determine radiographic signatures predictive of survival outcomes across tumor grades. We analyzed T1 post-contrast MRI sequences for 111 patients diagnosed with high or low grade glioma within The Cancer Genome Atlas. Our training cohort consisted of 80 gliomas (20 high grade, 60 low grade). An 8-layer 3D convolutional neural network (3D CNN) was constructed to classify low grade versus high grade gliomas. Stacked denoising auto-encoders were used for pre-training to initialize weights within the 3D CNN. Data augmentation was implemented to increase training sample size. Regularization and dropout were employed to prevent overfitting of the data. Our blinded validation cohort consisted of 31 gliomas. Performance of the model was evaluated using AUC and accuracy. The relationship between class predictions generated by the 3D CNN output and survival was studied using Kaplan Meier curves and the log rank test. Within our training cohort the 3D CNN achieved an accuracy of 95.0% and AUC of 0.90 (95% CI: 0.81-0.99). When tested on the validation cohort the 3D CNN achieved an accuracy of 80.6% and AUC of 0.81 (95% CI 0.69-0.92). Among low grade gliomas those with a high-grade glioma class prediction >30% were associated with a poorer overall survival (p<.001). Among high grade gliomas those with a low-grade glioma class prediction >25% were associated with an improved overall survival (p<.001). We developed a 3D CNN to accurately predict histopathologic grade based on MRI imaging features in patients with high and low grade gliomas. Additionally, we identified radiographic signatures predictive of survival across grade groups in our analysis, suggesting that radiographic features can be used in conjunction with clinical and molecular information to classify gliomas beyond histopathological grade.