Author: Steve Braunstein, Angela Jakary, Hui Lin, Bo Liu, Janine Lupo, Tiffany Ngan, Ke Sheng, Nate Tran 👨🔬
Affiliation: Radiation Oncology, University of California San Francisco, Graduate Program in Bioengineering, University of California San Francisco-UC Berkeley, Department of Radiation Oncology, University of California San Francisco, Department of Radiology and Biomedical Imaging, University of California San Francisco, Department of Radiation Oncology, University of California, San Francisco 🌍
Purpose: Current RT clinical target volumes (CTVs) for Glioblastoma (GBM) employ 2cm isotropic expansions of gross tumor volumes. However, studies showed patients still experience progression beyond these boundaries, underscoring the limitations of this approach in capturing GBM’s infiltrative nature. We hypothesize that integrating bio-clinical features with Density-Weighted White-Matter Path Length (DW-WMPL) mapping from diffusion MRI using a Cross-Modality Model (CMM) can better identify progression-prone regions, enabling patient-specific CTVs to enhance RT efficacy and outcomes.
Methods: We retrospectively analyzed longitudinal MRI data from 125 GBM patients at pre-surgery, post-surgery, and suspected recurrence per RANO guidelines. We incorporated imaging (T2-FLAIR, T1-weighted IR-SPGR, and DW-WMPL maps generated from diffusion tensor imaging) to quantify tumor infiltration probabilities beyond resection margins and along white-matter tracts. Tumor progression masks were auto-segmented via an in-house deep-learning model and verified by clinicians. The CMM was developed as a customized 3D-SwinUNETR incorporating a Bio2Image cross-attention block, integrating bio-clinical embeddings encompassing patient age, sex, extent of resection, and key molecular markers into encoded imaging features before decoding. We trained the CMM using Dice and Tversky loss, evaluating its performance on 25 hold-out cases with sensitivity, specificity, Dice score, Tversky coefficient, sparing index, and coverage index.
Results: Our proposed CMM outperformed standard RT planning, achieving a superior Dice score (0.64 ± 0.11) and Tversky coefficient (0.69 ± 0.12), with a 36% improvement in progression lesion coverage (p<0.001). Predicted CTVs significantly reduced volumes by 186 cc (p<0.0001), minimizing exposure to normal brain tissues. Compared to anatomical image-based models, CMM improved the Dice by 5%, Tversky by 11%, and progression coverage by 55%.
Conclusion: Cross-modality integration of bio-clinical and multi-parametric imaging features using a transformer-based model enhances identification of progression-prone regions while sparing normal tissues, paving the way for personalized RT planning.