Train Forwards, Optimize Backwards: Neural Surrogates for Personalized Medical Simulations

Jonas Weidner, Ivan Ezhov, Michal Balcerak, Bjoern Menze, Daniel Rueckert, Benedikt Wiestler

Keywords: Biophysical Tumor Modeling, Neural Surrogates, Gradient-Based Inverse Optimization

TL;DR: We build a differentiable system that personalizes biophysical tumor models by replacing slow PDE solvers with fast neural surrogates for gradient-based optimization.

Abstract: Personalizing biophysical simulations to individual patients remains a major computational bottleneck, as traditional optimization requires repeated runs of costly numerical solvers. We present a differentiable system that replaces the forward simulation with a neural surrogate, providing a fast and accurate approximation of the underlying biophysical model. The surrogate’s differentiability enables efficient gradient-based inversion of patient-specific parameters, even when the original solver is non-differentiable. Applied to a 3D finite-difference model of brain tumor growth, our method achieves clinically relevant accuracy while reducing optimization time from days to seconds. This demonstrates how differentiable surrogates can serve as core components of broader differentiable systems for scientific machine learning.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 41