Deep learning synthesis of Restriction Spectrum Imaging cellularity maps from conventional MRI in post-treatment glioblastoma.

objectiveThe feasibility of synthesizing Restriction Spectrum Imaging (RSI) cellularity maps from routinely acquired post-treatment MRI in glioblastoma was evaluated using a deep learning framework. MATERIALS AND

methodsA SwinUNETR model was trained on post-contrast T1-weighted, FLAIR, and ADC images from 184 imaging timepoints of 136 patients in the publicly available UCSD-PTGBM dataset, using five-fold patient-level cross-validation. Synthesis performance was assessed using tumor-ROI Pearson correlation as the primary endpoint, alongside secondary clinical agreement, hotspot localization, and scanner-stratified analyses.

resultsMean per-visit tumor-ROI Pearson correlation was 0.635 ± 0.164 (95% CI 0.608-0.662) and voxel-level intratumoral rank-agreement AUC for high-cellularity voxels was 0.800 (95% CI 0.785-0.814). SwinUNETR outperformed a 3D U-Net baseline across prespecified agreement and error metrics (all p < 0.001). A systematic negative bias of - 0.058 (95% limits-of-agreement width: 0.428) was observed within the tumor ROI, and hotspot Dice was 0.155 (95% CI 0.144-0.166), indicating limited focal localization. No significant difference was observed between scanner models for primary agreement metrics. DISCUSSION: Conventional post-treatment MRI approximates average RSI-derived cellularity but not the high-cellularity subregions with spatial precision. The approach may support screening where dedicated RSI is unavailable but cannot substitute for acquired RSI in quantitative or treatment-guiding applications.