RGNN: radiomic graph neural network for glioma grading utilizing 3d magnetic resonance images with clinical significant features

Abstract

Glioma, a prevalent and devastating brain tumor, presents formidable challenges in diagnosis and prognostication. This study endeavors to enhance glioma grading accuracy by leveraging 3D MRI data and a comprehensive array of medical features extracted through the PyRadiomic framework. Given the diverse manifestations of glioma tumors and their profound impact, an advanced approach is imperative for precise grading. This investigation meticulously extracts six distinct medical features, including First Order Statistics, Shape-based (3D), Gray Level Co-occurrence Matrix, Gray Level Run Length Matrix, Gray Level Size Zone Matrix, Neighboring Gray Tone Difference Matrix, and Gray Level Dependence Matrix. These features, computed based on tumor annotation, provide detailed characterizations of glioma tumors, elucidating their intricacies. To augment glioma grading accuracy further, various machine learning algorithms are employed. A pivotal contribution is the introduction of the Radiomic Graph Neural Network (RGNN) model, tailored for graph-based data, where nodes symbolize entities, and edges denote intricate relationships between them. The core objective of the RGNN model is to generate low-dimensional vector representations (embeddings) for nodes within the graph, preserving underlying structural and relational information. This innovative RGNN model significantly enhances precision in differentiating between various glioma grades. Specifically, for the Native T1 stage of MRI and T2-weighted (T2) stages, the proposed RGNN model achieves an unprecedented accuracy of 99.00%. This outperforms existing methods and sets a new benchmark in glioma tumor grading based on medical features, leveraging 4 stages of 3D magnetic resonance imaging.