T2-FLAIR digital subtraction radiomics versus neuroradiologist visual assessment for differentiating IDH-mutant astrocytomas from other non-enhancing low-grade gliomas: An externally validated machine learning study.

purposeNon-invasive differentiation of isocitrate dehydrogenase (IDH)-mutant, 1p/19q non-codeleted astrocytomas from other non-enhancing low-grade gliomas (LGGs) is crucial for treatment planning and prognostication, as these molecular subtypes have distinct therapeutic strategies and clinical outcomes. The conventional T2-weighted fluid-attenuated inversion recovery (T2-FLAIR) mismatch sign offers high specificity but limited sensitivity for this purpose. This study aimed to develop and externally validate a radiomics-based machine learning model using T2-FLAIR digital subtraction images to improve molecular subtype differentiation in non-enhancing LGGs and to compare its diagnostic performance with conventional neuroradiologist visual assessment.

methodsA total of 193 patients with non-enhancing LGGs were included from two independent cohorts: the Erasmus Glioma Database (EGD, n = 155, training) and The Cancer Genome Atlas Low-Grade Glioma (TCGA-LGG, n = 38, external test set). Histopathologically confirmed molecular subtypes included IDH-mutant, 1p/19q non-codeleted astrocytoma (positive class, n = 88) and other LGG subtypes comprising IDH-mutant, 1p/19q co-deleted oligodendroglioma and IDH-wildtype diffuse glioma (negative class, n = 105). T2-FLAIR digital subtraction images were generated by voxel-wise subtraction of co-registered FLAIR from T2-weighted images. Ten consensus radiomics features were selected using Least Absolute Shrinkage and Selection Operator (LASSO), minimum Redundancy Maximum Relevance (mRMR), and Boruta methods. An automated machine learning (AutoML) ensemble model was trained with 5-fold cross-validation. Two neuroradiologists (8 and 12 years of experience) independently assessed the conventional T2-FLAIR mismatch sign for comparison. Diagnostic performance was compared using the DeLong test.

resultsThe radiomics model achieved an area under the receiver operating characteristic curve (AUC) of 0.879 (95% confidence interval [CI]: 0.821-0.937; sensitivity: 71.2%, specificity: 85.4%) in the training cohort and 0.849 (95% CI: 0.741-0.957; sensitivity: 86.4%, specificity: 75.0%) in the external test set. The model outperformed conventional T2-FLAIR mismatch assessment by neuroradiologists (training AUC: 0.768, p = 0.003; external test AUC: 0.741, p = 0.038; DeLong test). The AUC difference of 0.030 between training and external test cohorts suggested adequate generalizability. Inter-reader agreement for conventional assessment was excellent (intraclass correlation coefficient = 0.89).

conclusionThe proposed T2-FLAIR digital subtraction radiomics model demonstrated superior diagnostic performance compared with conventional neuroradiologist visual assessment for differentiating IDH-mutant astrocytomas from other non-enhancing LGGs, primarily by improving sensitivity while maintaining acceptable specificity. This automated approach may serve as a complementary tool for non-invasive molecular characterization of diffuse gliomas, particularly in settings where subspecialty neuroradiology expertise may be limited.