Abstract
Glioma is a common malignant brain tumor that shows distinct survival among patients. The isocitrate dehydrogenase (IDH) gene mutation status provides critical diagnostic and prognostic value for glioma and is now accepted as the standard of care. A non-invasive prediction of IDH mutation based on the pre-treatment MRI has crucial clinical significance. Machine learning and deep learning models show reasonable performance in predicting IDH mutation status. However, most models neglect the systematic brain alterations caused by tumor invasion, where the infiltration along white matter tracts throughout the brain is identified as a hallmark of glioma. Structural brain network provides an effective tool to characterise brain organisation, which could be captured by the graph neural networks (GNN) for a more accurate prediction of IDH mutation status. Here we propose a method to predict the IDH mutation using GNN, based on the structural brain network of patients. Specifically, we firstly construct a network template of healthy subjects, which consists of atlases of edges (white matter tracts) and nodes (cortical and subcortical brain regions) to provide regions of interest (ROI). Next, we employ autoencoders to extract the latent multi-modal MRI features from the ROIs of the edge and node in patients. These features of edge and node of brain networks are used to train a GNN architecture in predicting IDH mutation status. The results show that the proposed method outperforms the baseline models using 3D-CNN and 3D-DenseNet. In addition, the model interpretation suggests its ability to identify the tracts infiltrated by tumor and corresponds to clinical prior knowledge. In conclusion, integrating brain networks with GNN offers a new avenue to study brain lesions using computational neuroscience and computer vision approaches.
Abstract (translated)
URL
https://arxiv.org/abs/2109.01854