An Explainable Machine Learning Model for Differentiation of Glioma Sub-types using MR Image Texture Analysis of Cerebral Edema

Abstract

Listen

Translate article

In this study, an attempt has been made to develop an explainable diagnostic model for glioma sub-types using texture features extracted from Magnetic Resonance (MR) images of cerebral edema. For this, structural MR brain images of Low-Grade Glioma (LGG) and High-Grade Glioma (HGG) subjects are considered from a public database. The corresponding edema sub-regions are also obtained from the database. Texture features are extracted from edema sub-regions and machine learning classifiers namely decision trees, random forest and eXtreme Gradient Boosting are employed to differentiate LGG and HGG. Further, SHapley Additive exPlanations (SHAP) are used to explain the model output. Results indicate that the developed model is able to identify glioma. Random forest classifier achieves the highest area under receiver operating characteristic curve of 0.79 in differentiating LGG and HGG. Among the extracted features, SHAP values suggest that the texture descriptors from gray level co-occurrence matrix is highly associated with the classification of glioma sub-types. Thus, the proposed MR image-based machine learning model focusing on the cerebral edema aid in the automated diagnosis of glioma.