UCSF

Abstract Recent studies have highlighted the importance of using molecular biomarkers (IDH, 1p19q, TERT) to group gliomas that have similar clinical behavior, response to therapy, and outcome. An emerging hypothesis is that glioma specific genetic and/or molecular alterations manifest as specific observable changes in MR anatomical imaging. Morphologic and texture features, originating from anatomical MRI, have been investigated as imaging biomarkers to predict MGMT and glioma group status. These texture or morphologic based approaches pose several challenges including requirements for several preprocessing steps such as intensity standardization, skull stripping, and tumor segmentation. Deep learning is an important evolving technology in many different fields, including anatomic imaging, and can be used to empirically identify important features in a variety of modalities, including MRI. Importantly deep learning precludes the need for extensive pre-processing. We describe a convolutional neural network, evaluating resnet50, vgg16, inception and xception neural network architectures, that can predict IDH, 1p19q and TERT status utilizing conventional T2 weighted MRI imaging with intensity normalization and nonuniform intensity normalization (N4) bias corrections. The dataset consisted of 432 images (340 for training and 92 for validation) from patients published by Eckel-Passow et al New England Journal of Medicine (2015). The system achieved a weighted f1 score of 0.901, 0.937 and 0.924 for IDH, 1p19q and TERT prediction on the test dataset, respectively. IDH status was misclassified in 9 out of 92 patients, while 1p19q and TERT status was misclassified in 6 and 7 patients respectively. Our results demonstrate the potential of deep learning architectures applied to conventional MRI to predict molecular glioma groups.