IMG-134. AI-based multimodal integration of clinical features with white matter infiltration maps for personalized RT planning of glioblastoma

PURPOSE Current RT clinical target volumes (CTVs) for glioblastoma employ a 2cm isotropic expansion of gross tumor volumes. However, studies showed patients still experience progression beyond these boundaries, underscoring the limitations of this approach in capturing GBM’s anisotropic infiltrative nature. This study developed and clinically validated a cross-modality deep learning model (CMM) integrating bio-clinical features with Density-Weighted White-Matter Path Length (DW-WMPL) mapping to predict GBM progression and come up with CTVs that better capture progression-prone regions compared to standard RT CTVs. METHODS We retrospectively analyzed longitudinal MRI from 125 GBM patients at pre-surgery, post-surgery, and recurrence per RANO guidelines. The CMM was developed as a customized 3D-SwinUNETR architecture incorporating a Bio2Image cross-attention block that integrated bio-clinical embeddings encompassing patient age, sex, extent of resection, and key molecular markers (MGMT, PTEN, and EGFR) into encoded imaging features. Tumor progression masks were auto-segmented and clinically verified. We trained the CMM using Dice and Tversky loss and 5-fold cross-validation, evaluating the performance of the ensembled model on 25 hold-out cases with Dice, Tversky coefficient, sparing index, and coverage index. For clinical validation, we generated comparative RT plans for 12 holdout patients using conventional CTVs versus DW-WMPL-enhanced CTVs, both delivering 60Gy while meeting NRG dose constraints. RESULTS Our CMM outperformed standard T2-lesion+2cm expansion, achieving a superior Dice score (0.716 ± 0.084) and Tversky coefficient (0.755 ± 0.126), with a 25% improvement in progression lesion coverage (p<0.001), while reducing CTV volumes by 186 cc (p<0.0001), RT plans using DW-WMPL-enhanced CTVs significantly improved median prescription coverage of actual disease progression by 21.5% (67.5% vs. 88.9%, p=0.034) compared to conventional CTVs, with numerically lower maximum doses to brainstem, optic chiasm, and optic nerves. CONCLUSION Cross-modality integration of bio-clinical and multi-parametric imaging features using a transformer-based model improves both progression prediction accuracy and RT planning outcomes, providing better coverage of recurrence-prone regions while sparing critical structures and normal brain.