The explainability of radiomic-based machine learning models for brain glioma grading on amide proton transfer-weighted images

This study aimed to evaluate radiomic-based machine learning models for glioma grading on amide proton transfer weighted (APTw) images using explainability algorithms. A total of 102 patients who underwent preoperative MR examinations, including FLAIR, T1-weighted, T1-weighted contrast-enhanced, and APTw images, were included. Two groups of APTw images were analyzed: one corresponding to contrast-enhanced regions of gliomas and the other corresponding to both contrast-enhanced and peritumoral edematous regions of gliomas. Radiomic features were extracted from these regions. Random forest, support vector machine, naïve bayes classifier, and logistic regression models were trained to distinguish grade 4 from non-grade 4 gliomas. These models were analyzed by Shapley values, permutation importance, and the method of anchors. The results of model explainability analysis revealed that the grading performance of these models relied on radiomic features highlighting the heterogeneity of radiologic phenotypes.