10.1371/journal.pmed.0020297

Vollständiger Abstract

Worum geht es in dieser Arbeit?

This study aimed to develop and externally validate a radiomics-based machine learning framework for the noninvasive differentiation of non-enhanced glioblastoma (NE-GBM) from astrocytoma, IDH-mutant, grade 2 (IDHm-A2), thereby addressing the diagnostic limitations of qualitative MRI and the challenge of data scarcity. This diagnostic study used a radiomics-based machine learning framework, with integrated histopathological and molecular diagnosis serving as the reference standard. We retrospectively enrolled 84 glioma patients (July 2021-November 2022) from our center, comprising 34 pathologically confirmed NE‑GBM. Radiomic features were extracted from T1, T2, FLAIR, and ADC sequences. The framework was developed to differentiate NE‑GBM from IDHm‑A2 and was subsequently validated on an internal test cohort (n = 22) and an external test cohort (n = 12). Diagnostic performance was assessed using ROC analysis and decision curves, and model predictions were interpreted via SHapley Additive exPlanations (SHAP). A total of 118 patients (47 NE‑GBM, 71 IDHm‑A2) were evaluated. Fifty‑four radiomic features and age were selected for model construction. The optimized model demonstrated strong discriminatory performance, with the internal test cohort achieving an AUC of 0.935 and an accuracy of 91.0%, and the external test cohort yielding an AUC of 0.969 and an accuracy of 90.0%. The proposed framework provides an accurate, noninvasive tool for distinguishing NE‑GBM from IDHm‑A2, directly addressing this critical diagnostic challenge.

Abstract: PubMed · Datensatz