Abstract
Diffuse gliomas are malignant brain tumors that grow widespread through thebrain. The complex interactions between neoplastic cells and normal tissue, aswell as the treatment-induced changes often encountered, make glioma tumorgrowth modeling challenging. In this paper, we present a novel end-to-endnetwork capable of future predictions of tumor masks and multi-parametricmagnetic resonance images (MRI) of how the tumor will look at any future timepoints for different treatment plans. Our approach is based on cutting-edgediffusion probabilistic models and deep-segmentation neural networks. Weincluded sequential multi-parametric MRI and treatment information asconditioning inputs to guide the generative diffusion process as well as ajoint segmentation process. This allows for tumor growth estimates andrealistic MRI generation at any given treatment and time point. We trained themodel using real-world postoperative longitudinal MRI data with glioma tumorgrowth trajectories represented as tumor segmentation maps over time. The modeldemonstrates promising performance across various tasks, including generatinghigh-quality multi-parametric MRI with tumor masks, performing time-seriestumor segmentations, and providing uncertainty estimates. Combined with thetreatment-aware generated MRI, the tumor growth predictions with uncertaintyestimates can provide useful information for clinical decision-making.