Deep learning-based quantification of T2-FLAIR mismatch sign: extending IDH mutation prediction in adult-type diffuse lower-grade glioma

Abstract

Objectives: To investigate the predictive value of the quantitative T2-FLAIR mismatch ratio (qT2FM) with fully automated tumor segmentation in adult-type diffuse lower-grade gliomas (LGGs). Materials and methods: This retrospective study included 218 consecutive patients (mean age, 47 years ± 15 [SD]; 125 males) diagnosed with adult-type diffuse LGG. The cohort was classified into IDH wild-type (IDHwt), IDH-mutant with 1p/19q-codeletion (IDHmut-Codel), and IDH-mutant without 1p/19q-codeletion (IDHmut-Noncodel) subtypes. Tumor masks were obtained using deep learning-based segmentation, and qT2FM was calculated from the differences in signal intensity ratios on T2 and FLAIR images. Multivariable logistic regression identified predictors for identifying IDHmut-Noncodel and IDH mutation status. Point-biserial correlations were analyzed between qualitative and quantitative T2FM, and median apparent diffusion coefficient (ADC) value. Diagnostic performance was evaluated with a receiver operating characteristic curve. Results: The IDHmut-Noncodel group had a higher qT2FM (0.37 ± 0.38, p = 0.004) than the IDHmut-Codel (0.24 ± 0.39) and IDHwt groups (0.07 ± 0.62). The qT2FM was the only independent imaging predictor for identifying IDHmut-Noncodel (OR = 3.43, 95% CI: 1.30–9.05, p = 0.01). Independent predictors of IDH mutation were younger age (p < 0.001), frontal lobe location (p = 0.007), cortical involvement (p < 0.001), and higher qT2FM (p = 0.034). The qT2FM significantly correlated with visual T2FM (vT2FM) and median ADC value. Adding qT2FM to vT2FM improved performance in identifying IDHmut-Noncodel (AUC 0.77, 95% CI: 0.70–0.82) and IDH mutation status (AUC 0.77, 95% CI: 0.71–0.83) than each parameter alone. Conclusion: The qT2FM ratio, derived from deep learning-based tumor segmentation, is a valuable predictor for identifying IDH mutation status and the IDHmut-Noncodel subtype in patients with adult-type diffuse LGG. Key Points: Question Does deep-learning-based quantification of the T2-FLAIR mismatch sign provide accurate prediction of IDH-mutant, 1p/19q non-codeleted astrocytomas and enhance identification of IDH mutation status? Findings Quantifying the T2-FLAIR mismatch sign with a fully automated segmentation tool achieved high accuracy in identifying IDH-mutant, 1p/19q non-codeleted astrocytomas, and enhanced IDH status prediction. Clinical relevance Integrating the qT2FM into clinical protocols enhances diagnostic precision and guides treatment strategies, underscoring the role of advanced imaging in neuro-oncology.