Amide proton transfer and arterial spin labeling for non-invasive molecular stratification of glioma: a multi-dataset imaging biomarker study

To evaluate amide proton transfer chemical exchange saturation transfer magnetization transfer ratio asymmetry (MTRasym) and arterial spin labeling cerebral blood flow (CBF), two advanced MRI-based maps, for predicting glioma isocitrate dehydrogenase (IDH) mutation status, 1p/19q codeletion status, and grade using radiomics features; to assess cross-dataset/cross-vendor generalizability; and to examine whether combining MTRasym and CBF features improves prediction. This multi-center study included 219 grade 2–4 glioma patients from three datasets (Netherlands/D1: n = 48, Siemens, prospective; Russia/D2: n = 42, Philips, retrospective; China/D3: n = 129, Siemens, retrospective). Descriptive and first-order radiomic features were analyzed from MTRasym (n = 219) and CBF maps (n = 90, CBF not available in D3), with and without normalization to contralateral regions of interest (ROI). Univariate diagnostic performance was assessed using area under the receiver operating characteristic (AUC) curves, random forest radiomics classifiers were evaluated using 5-fold cross-validation on D1 and subsequently validated externally on D2/D3. Combined MTRasym+CBF models were assessed on D1/D2 (n = 90). Tumor-to-contralateral ROI-normalized ratios were comparable across sites (p > 0.05). Univariate MTRasym achieved AUCs of 0.94 (1p/19q), 0.87 (grade), and 0.76 (IDH); contralateral ROI-normalized CBF performed comparably for IDH (AUC = 0.77) and 1p/19q (AUC = 0.75), but not grade (AUC = 0.72). Radiomics models achieved AUCs of 0.91 (1p/19q), 0.81 (IDH), and 0.89 (grade) for MTRasym; combined MTRasym+CBF models significantly improved AUCs of 0.85 (IDH, p = 0.007) and 0.91 (grade, p = 0.001). External validation revealed MTRasym performed well on D3 (Siemens, same vendor), while normalized CBF generalized better than MTRasym to D2 (Philips). MTRasym and CBF features show promise for glioma molecular stratification, and multimodal modeling improves predictions. Cross-vendor validation reveals modality-specific generalizability patterns.