Defining Radiation Target Volumes with AI-Driven Predictions of Glioma Recurrence from MRSI, Diffusion MRI, and Transformers

Motivation: Presurgery MR scans have a higher percentage of tumor voxels and can be used as a better signal to predict tumor progression using AI-driven models. Goal(s): We show deep learning can be used to predict tumor progression in patients diagnosed with glioblastoma multiforme (GBM) using a combination of anatomical, diffusion, and metabolic MRI scans done prior to surgery. Approach: Convolutional Neural Networks (CNNs) and Vision Transformers are trained to predict tumor ROI of the progression lesion using presurgery MR scans. Results: Our methods perform better than standard of care in both inclusion of the tumor and exclusion of the normal brain. Impact: Our results highlight the potential value of deep learning in future RT treatment planning with presurgery MRI scans. Vision transformers perform at par (if not better) with CNNs suggesting opportunities for future work into their use in progression prediction.