Deep Generative Models for the Prediction of Brain Tumor Progression from Multimodal Data

Deep learning has revolutionized medical imaging, enabling unprecedented accuracy in tasks such as organ segmentation, lesion detection, and tissue classification. It becomes increasingly important to apply these methods to complex clinical predictions, such as prediction of cancer progression and treatment response. In this thesis, we focus on predicting spatial pixel-wise brain tumor evolution, particularly glioma progression, by developing a novel framework for spatial modeling of tumor growth. Such prediction tasks are inherently complex and rely on vast, multimodal, longitudinal datasets to capture the numerous factors influencing tumor behavior. However, medical imaging data are chronically scarce due to burdensome acquisition protocols, privacy regulations, and the limited survival time of glioma patients, further exacerbated by the irregularly acquired longitudinal MRI data in clinical practice, leading to incomplete follow-up sequences and data imbalances. To address these challenges, we propose in the first part of our work two generative augmentation method, leveraging variational autoencoders and diffusion-based models to synthesize high-fidelity longitudinal medical images and their corresponding segmentation masks, providing a clear spatial representation of generated tumors. The augmented datasets substantially improve training for downstream prediction models. Building on these enriched datasets, we propose in the second part of our work, an end-to-end multitask diffusion framework that integrates synthetic medical images and uncertainty quantification to forecast spatial tumor progression from as few as two early MRI timepoints. Our model accurately predicts future tumor extent up to several months ahead, achieving a mean Dice similarity coefficient of 0.75 on held-out public and private test sets, and demonstrating robust generalization across imaging centers. By combining generative augmentation approaches with a multi-task diffusion-based predictor, we show the possibility of accurate, time-agnostic glioma progression modeling from limited longitudinal MRI data. This work represents a significant step toward improving clinical decision-making in neuro-oncology, enabling personalized treatment planning and better patient outcomes.