Graph-Radiomics Learning (GrRAiL): Application to Distinguishing Glioblastoma Recurrence from Pseudo-Progression on Structural MRI

A significant challenge in managing aggressive cancers is the ability to quantify the intra-tumoral heterogeneity that contributes to poor prognostic outcomes. While radiomics approaches have previously attempted to characterize tumor heterogeneity, they have primarily focused on utilizing textural- and morphological-based features from the tumor confines on imaging. In this work, we seek to go beyond textural intensity patterns and present a new Graph-Radiomics Learning (GrRAiL) approach that utilizes graph patterns to capture the spatial interactions in radiomic features. The rationale for this work is that the spatial interactions in radiomic expressions when mapped via connected graphs, can model the extensive heterogeneity of the tumor. GrRAiL involves (a) extracting radiomic expression maps from the segmented lesion, (b) clustering the sub-regions within the expression maps, (c) graph construction and feature extraction, followed by using machine learning approaches to quantify the tumor regions. We assess the efficacy of GrRAiL in accurately quantifying lesion heterogeneity to address a challenging clinical dilemma in Glioblastoma (GB), the most aggressive brain tumor. In this context, we seek to utilize GrRAiL to differentiate GB true progression (TP) from pseudo-progression (PsP), a benign radiation-induced treatment effect that often mimics the appearance of TP on MRI scans. Our study included n = 106 GB patients from Cleveland Clinic (TP: 38, PsP: 22 for training) and Dana-Farber Cancer Center (TP: 33, PsP: 13 for testing). After extracting GrRAiL features, a Random Forest classifier was employed to distinguish TP from PsP. Further, GrRAiL’s performance was compared with radiomic features as well as intensity-based graph features, in distinguishing TP from PsP. GrRAiL achieved a test accuracy of 0.76 and outperformed radiomics (accuracy = 0.69) and intensity-based graph approaches (accuracy = 0.61).