Multiparametric Quantitative MRI With 2.5D Deep Learning for Molecular Classification and Overall Survival Assessment in Adult-Type Diffuse Gliomas.
Authors
Affiliations (9)
Affiliations (9)
- Department of Radiology, Tangdu Hospital, The Fourth Military Medical University, Xi'an, Shaanxi, China.
- Functional and Molecular Imaging Key Lab of Shaanxi Province, Xi'an, Shaanxi, China.
- 7T MRI Precision Neurology Platform of Shaanxi Province, Xi'an, Shaanxi, China.
- Department of Rehabilitation, Tangdu Hospital, The Fourth Military Medical University, Xi'an, Shaanxi, China.
- MR Research, GE Healthcare, Beijing, China.
- Department of Radiology, General Hospital of Ningxia Medical University, Yinchuan, Ningxia, China.
- Department of Magnetic Resonance, Lanzhou University Second Hospital, Lanzhou, Gansu, China.
- Gansu Province Clinical Research Center for Functional and Molecular Imaging, Lanzhou, Gansu, China.
- Gansu Province Medical MRI Equipment Application Industry Technology Center, Lanzhou, Gansu, China.
Abstract
Quantitative MRI may enable noninvasive glioma characterization, but optimal analytic strategies and prognostic utility remain uncertain. To compare volume-of-interest (VOI)-mean, radiomics, and 2D, 2.5D, and 3D deep learning (DL) for preoperative molecular classification and assess the prognostic value of a 2.5D-DL-derived isocitrate dehydrogenase wild-type score (2.5D-IDHwt score) in a postoperative clinicopathologic survival model. Retrospective multicenter study. One thousand four hundred and eight adults with diffuse glioma: training (n = 409), temporal validation (n = 300), and three external test cohorts (n = 119, 310, and 270). 3 T; spin-echo echo-planar diffusion-weighted imaging, 3D pseudo-continuous arterial spin labeling, and multidynamic multiecho synthetic MRI. Apparent diffusion coefficient, cerebral blood flow, T1, T2, proton-density, and contrast-enhanced T1 maps were generated. Two neuroradiologists (5 and 20 years' experience) delineated whole-tumor VOIs. Six map-specific models were stacked separately for VOI-mean, radiomics, 2D-DL, 2.5D-DL, and 3D-DL. 2.5D-qMR-Clin combined the 2.5D-IDHwt score with age, O<sup>6</sup>-methylguanine-DNA methyltransferase promoter methylation, and extent of resection and was compared with PathoClin incorporating pathological IDH status. Area under the curve (AUC), DeLong tests, Harrell C-index, bootstrap resampling, calibration, decision-curve analysis, Kaplan-Meier analysis, and Cox regression. p < 0.05 indicated significance. The 2.5D-DL stacking model achieved the highest external AUCs for IDH mutation (0.89-0.93) and 1p/19q codeletion (0.92-0.95), significantly outperforming the other four stacking models. Reverse Kaplan-Meier median follow-up was 1344 days (95% CI, 1316-1383); 27.5% died. External C-indices for 2.5D-qMR-Clin were 0.750, 0.778, and 0.762 versus 0.704, 0.722, and 0.723 for PathoClin, with significant differences in test cohorts 2 and 3. Fixed-cutoff risk stratification was significant across all external cohorts (hazard ratios, 5.30-5.80). A web-based 2.5D-qMR-Clin explorer provided 1-, 2-, and 3-year OS estimates. Among evaluated approaches, 2.5D-DL stacking performed best for molecular classification, and the 2.5D-IDHwt score improved OS assessment in the postoperative clinicopathologic model. 3. Stage 2.