Enhanced Glioma Genotype Prediction Using CEST MRI With Full Z-Spectrum Input, Pixel-Level Learning, and Majority Voting.

Abstract

Accurate glioma genotype prediction, such as the isocitrate dehydrogenase (IDH) mutation and O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status, is critical for personalized treatment planning and prognosis. Chemical exchange saturation transfer (CEST) MRI enables noninvasive assessment of tumor metabolism and the microenvironment, offering promising potential for genotype prediction. However, existing CEST-based diagnostic methods are limited by information loss from simplified quantification and by instability arising from ROI-dependent analysis. To address these limitations, a deep learning framework was developed that integrates the full Z-spectrum input, pixel-level training, and a majority voting mechanism to improve prediction accuracy and robustness. In a cohort of 84 patients, a feedforward neural network (FNN) was trained to process the entire Z-spectrum and generate pixel-wise predictions, which were subsequently aggregated into patient-level outcomes through majority voting. T-distributed stochastic neighbor embedding (t-SNE) was applied to evaluate the effectiveness of the full Z-spectrum input. Model generalizability and robustness were assessed using fivefold cross-validation, and stability was quantified via the coefficient of variation (CoV). Prediction performance was evaluated in terms of accuracy, sensitivity, specificity, and the area under the curve (AUC). The proposed method demonstrated improved interclass separability with the full Z-spectrum input (t-SNE score: 64.50), outperforming APTw (55.60) and Lorentzian fitting (56.97). Stable prediction performance was achieved for both IDH (accuracy: 0.86 ± 0.04 and AUC: 0.91 ± 0.03) and MGMT (accuracy: 0.82 ± 0.02 and AUC: 0.91 ± 0.04) genotypes. The approach also showed strong robustness to ROI selection, with a CoV of 0.69%, compared to 7.44% for APTw-mean and 3.67% for Lorentzian fitting (LF-mean). These findings support the effectiveness of combining the full Z-spectrum input, pixel-level learning, and majority voting to improve the reliability of noninvasive glioma genotype prediction using CEST MRI.