IMG-04. Anatomical tumor growth predictions in pediatric diffuse midline glioma using generative AI

Abstract Pediatric diffuse midline glioma (DMG) stands out among other tumor subtypes with a particularly poor prognosis. Standardized radiotherapy remains the key palliative care for these patients. Personalized geometric dose distributions have the potential to maximize the benefit for the individual. We aimed to demonstrate the feasibility of novel computer vision approaches for generating clinically relevant predictions of anatomical tumor growth. Magnetic resonance imaging (MRI) for adult and pediatric brain tumor patients from the BraTS23 challenge were used to train a 2D guided denoising diffusion implicit model (DDIM). MRIs with enlarged tumor sizes were generated from baseline images and assessed quantitatively by Structural Similarity Index (SSIM) and statistical analysis of radiomic features. Classification of real and generated images by radiologists was performed as a qualitative evaluation. Predictions were validated on an independent dataset of longitudinal MRIs from five DMG patients. We generated high-quality scans, supported by a high SSIM (0.8) with non-significant differences in most radiomic features (83%). Expert radiologists failed to correctly distinguish generated from real MRIs (mean precision 0.55 and recall 0.41). Leveraging the probabilistic nature of DDIMs, we compiled tumor growth probability heatmaps through repeated generations which successfully translated to a 3D prediction space by applying it to consecutive 2D slices. The generated anatomical predictions closely resembled the observed tumor progression in the longitudinal pediatric DMG subset. Notably, the direction and extent of tumor growth align while preserving the anatomical characteristics of the brain, yielding a mean DICE score of 0.8 between predicted and true tumor. Further quantitative validation on larger longitudinal datasets is required, to ensure the method’s robustness. The proposed framework enables a personalized approach for defining radiation therapy target regions which could translate to improved clinical outcomes.