Gli-Ode: Encoder-Decoder System with Neural ODEs for Dynamic GBM Growth Forecasting

Overview

Glioblastoma multiforme (GBM) is the most lethal brain cancer due to its highly diffusive, irregular growth patterns, leaving recurrent tumor cell pockets that can remain undetected with medical imaging and treatment margins, allowing for tumor recurrence. However, by approximating spatiotemporal tumor concentrations with tumor modeling, treatment areas can be refined for recurrence-prone pockets during radiochemotherapy treatment. Recent deep learning approaches in tumor modeling have facilitated patient personalized therapy with competitive runtimes, but three limitations hinder clinical integration: (1) dependence on longitudinal pretreatment MRI image data, unavailable for most patients; (2) reliance on fixed-depth models, preventing arbitrary time forecasting; and (3) interpretability challenges in tumor growth approximations, arising from fixed-depth predictions. This study intends to overcome these flaws with Gli-Ode, a GBM growth modeling tool trained on 5000 synthetic tumor pairs (from day 50 to day 550) that provides continuous-time tumor concentration predictions based on a single tumor slice. Gli-Ode employs an encoder-decoder architecture where the encoder compresses the input slice into a latent feature vector. A Neural Ordinary Differential Equation (NeurODE) module then evolves this latent representation of tumor features over time, allowing the decoder to reconstruct the predicted tumor image at any desired time point, days 50-500. Gli-Ode can predict tumor evolution with high accuracy and adapt to arbitrary time steps, showcasing strong potential for clinical integration. Performance of Gli-Ode is measured by simulating 100 synthetic brain geometries (days 50-550) against their ground truth tumor concentrations and measured with (1) regression error rates (2) similarity metrics and (3) runtime comparisons with state-of-the-art numerical solvers and models. The results indicate that Gli-Ode can approximate tumor concentrations with high accuracy while retaining competitive runtimes, making it a viable tool for clinical use.