Patient-specific prediction of glioblastoma growth via reduced order modeling and neural networks

Cerrone et al. (Math Biosci 2025): diffuse-interface (Cahn–Hilliard-type) two-phase GBM growth model with oxygen reaction–diffusion and DTI-derived anisotropic tensors, reduced with a 20-function POD basis plus neural networks (POD-NN, ~150x faster than the full-order finite-element solver: 5 s vs 780 s); an inverse network estimates 6 patient-specific parameters from tumor distributions at two time points 20 days apart. The authors report ~96% tumor-volume forecasting accuracy and ~15% error on normalised parameters; Morris sensitivity: proliferation and oxygen-consumption rates dominate, oxygen supply moderate; a local Monte Carlo check (1% input perturbation, single reference parameter set) shows larger uncertainty for hypoxia threshold and brain stiffness. Limits: trained and validated on synthetic data only (one clinical case's anatomy used for the mesh); no treatment effects (real follow-up patients received Stupp therapy) and no phenotypic heterogeneity; clinical MRI follow-up is every two months, unlike the 20-day model interval; real-patient application is an ongoing study (GLIOMATH prospective observational trial).

Authors / organisation

Cerrone D, Riccobelli D, Gazzoni S, Vitullo P, Ballarin F, Falco J, Acerbi F, Manzoni A, Zunino P, Ciarletta P.

Venue

Mathematical Biosciences

Date

2025-06-03

Type

paper

Year

2025

DOI

10.1016/j.mbs.2025.109468

PMID

40473100

Verification

arXiv API 2412.05330 (journal_ref Mathematical Biosciences 387:109468 (2025); v1 2024-12-04, v2 2025-04-01) + DOI 10.1016/j.mbs.2025.109468 + full text read 2026-10-19; Crossref (title, 10 authors, vol 387, art. 109468, issue 2025-09, VOR licence start 2025-06-03) + PubMed PMID 40473100 (epub 2025 Jun 3, print 2025 Sep; title/journal match); date = online publication date

GBMDigital twinMechanistic & agent-based modelMRIreduced-order-modelneural-network

Open in the searchable library → · JSON