Rapid Personalization of PDE-Based Tumor Growth using a Differentiable Forward Model

Jonas Weidner, Ivan Ezhov, Michal Balcerak, bjoern menze, Benedikt Wiestler

Keywords: gradient-based optimization, PDE-based growth models, tumor modeling

Abstract: Partial differential equation (PDE) based brain tumor growth models have the potential to personalize glioma therapy. However, calibrating these models to individual patients is computationally expensive using traditional optimization techniques. In this work, we propose an approach leveraging the differentiability of deep learning (DL) based PDE forward solvers to efficiently calibrate the tumor models. Through gradient-based optimization with respect to the input tumor parameters, we iteratively minimize the loss between the predicted and actual tumor distribution in the patient's MRI scans. We evaluate our method on a cohort of nine glioma patients and demonstrate a dramatic reduction in the time to solve the inverse problem from hours, using typically employed evolutionary sampling or Monte Carlo methods, to minutes while achieving comparable modeling results.

Submission Number: 117