Radiomics and machine learning analysis to differentiate low grade glioma and encephalitis on non-contrast MRI

Accurate differentiation between low-grade gliomas and encephalitis remains challenging due to overlapping MRI features. This study investigated whether MRI-based radiomics derived from routine non-contrast MRI could improve the differentiation of these entities in diagnostically challenging cases presenting in emergency settings. This retrospective single-center study included 41 patients with ambiguous MRI findings who underwent 3T MRI within 48 h of symptom onset. Radiomic features extracted from T2-FLAIR, DWI, and T1-TSE images, together with laboratory data available at presentation, were analyzed. Random Forest classifiers were developed using selected radiomic features, and model performance was evaluated using nested cross-validation, independent validation, receiver operating characteristic (ROC) analysis, and DeLong’s test. The cohort included 23 patients with histologically proven low-grade gliomas and 18 with confirmed encephalitis and was divided into training ( n = 31) and validation ( n = 10) sets. FLAIR-based models showed the best discriminatory performance, achieving an AUC of 0.82 (95% CI: 0.54–1.00) for both the radiomics-only and radiomics-clinical models outperforming T1-weighted (AUC 0.52 for radiomics-only and 0.60 for radiomics-clinical model) and DWI models (AUC 0.72 for both radiomics-only and radiomics-clinical approaches). The combined-sequence radiomics-clinical model achieved an AUC of 0.80 (95% CI: 0.41–1.00), with 100% sensitivity, 80% specificity, and 90% accuracy. Comparison of the ROC curves using DeLong’s test did not demonstrate statistically significant differences between the radiomics-only and radiomics-clinical models for any MRI sequence. This proof-of-concept study suggests that radiomic features extracted from routine non-contrast MRI may help differentiate low-grade gliomas from encephalitis in diagnostically challenging cases; however, these preliminary findings warrant validation in larger multicenter cohorts before clinical implementation.