A random graph-based neural network approach to assess glioblastoma progression from perfusion MRI

Abstract

The identification of glioblastoma progression from the non-contrast enhancing areas is critically challenging to clinicians. This leads to poor prognosis and a very high local progression rate even after surgical resection and chemo-radiotherapy. Deep learning methods have been aiding in tracking tumor segmentation and progression. However, feature-based learning from these methods requires large datasets, multiple annotations, and measurements, resulting in complex, time-consuming networks. This work introduces a random Graph Neural Network (GNN) approach using two 16 × 16 random graph convolution layers to perform pixel-wise classification and predict glioblastoma progression from advanced perfusion MRI. A total of 17 patients, newly diagnosed with glioblastoma and treated with surgery and standard concomitant chemo-radiation therapy (CRT), were examined from The Cancer Imaging Archive Brain-Tumor-Progression dataset. Dynamic susceptibility contrast (DSC) MRI exams generated within 90 days following CRT completion and at clinically determined progression were evaluated for each patient. Three DSC modalities, such as normalized cerebral blood flow, normalized relative cerebral blood volume, and standardized relative cerebral blood volume were considered for the study. The proposed model provided an overall competitive accuracy of 99.76% with a training time of 6 minutes and a test time of 7.36 seconds. The proposed random GNN model demonstrates promising potential to predict final ground-truth maps accurately.