Individualizing Glioma Radiotherapy Planning by Optimization of Data and Physics-Informed Discrete Loss

Michal Balcerak, Jonas Weidner, Petr Karnakov, Ivan Ezhov, Sergey Litvinov, Petros Koumoutsakos, Tamaz Amiranashvili, Ray Zirui Zhang, John S. Lowengrub, Bene Wiestler, Bjoern Menze

TL;DR

The GliODIL framework personalizes radiotherapy for brain cancer patients by predicting invisible tumor infiltration using a physics-informed optimization approach. By treating biological growth equations as "soft constraints" rather than rigid rules, it outperforms standard clinical margins and traditional models in predicting where a tumor is likely to recur.

What Problem Does This Paper Address?

Gliomas, particularly glioblastomas, are notoriously difficult to treat because they don't just grow as a solid lump; they infiltrate the surrounding brain tissue like a microscopic fog. Current clinical practice relies on a "one-size-fits-all" approach: doctors draw a uniform 1.5 cm safety margin around the visible tumor on an MRI to target these hidden cells.

This approach is flawed. It often over-treats healthy brain tissue, causing unnecessary side effects, or under-treats areas where the tumor has spread further than expected along specific anatomical pathways. While mathematical models (like reaction-diffusion equations) exist to simulate this growth, they are often too slow for clinical use or too rigid to handle the messy, complex realities of individual patient data.

Methods

The authors introduce GliODIL (Glioma Optimizing a Discrete Loss), a framework that blends traditional numerical physics with modern optimization techniques.

  1. The Physics Model: It uses the Fisher-Kolmogorov (FK) equation, which models tumor growth through two main processes: cell migration (diffusion) and cell proliferation.
  2. Optimizing a Discrete Loss (ODIL): Unlike Physics-Informed Neural Networks (PINNs), which can be notoriously difficult to calibrate, GliODIL discretizes the PDE onto a multi-resolution grid. It then optimizes a loss function that balances three things:
  3. Physics Loss ($L_{PDE}$): How well the solution follows the FK equation.
  4. Data Loss ($L_{DATA}$): How well the solution matches the patient's MRI and FET-PET scans.
  5. Initial Condition Loss ($L_{IC}$): The assumption that the tumor started from a focal point.
  6. Soft Constraints: Crucially, GliODIL treats the physics as a "soft constraint." If the biological model can't perfectly explain the imaging data (which happens in complex, multi-focal cases), the optimizer can "loosen" its adherence to the PDE to better reflect the observed reality.
The GliODIL pipeline integrates multi-modal imaging with a physics-based growth model to generate personalized radiotherapy plans.

Datasets

  • Synthetic Dataset: 100 simulated patients (50 single-focal, 50 multi-focal) used to calibrate the weight of the physics loss ($\lambda_{PDE}$).
  • Clinical Dataset: 152 glioblastoma patients from TUM University Hospital. This includes 58 cases with both MRI and FET-PET (a metabolic imaging modality) and 94 cases with MRI only. The dataset includes post-treatment follow-up scans to track where tumors actually recurred.

Key Insights and Ideas

  • The "Goldilocks" Physics Weight: The researchers found that strictly adhering to the growth equation ($\lambda_{PDE} \gg 1$) is actually detrimental for complex tumors. By setting $\lambda_{PDE} = 1$, GliODIL achieves a balance that captures both standard growth patterns and unexpected infiltrations visible only in metabolic scans.
  • FET-PET is a Game Changer: Integrating metabolic data (FET-PET) significantly improved the model's ability to identify high-risk areas that structural MRI misses.
  • Computational Efficiency: By using a multi-resolution grid and automatic differentiation, GliODIL performs parameter inference in about 45 minutes on a single GPU—a significant improvement over traditional Bayesian methods that can take hours on large CPU clusters.

Key Results

The primary metric was Recurrence Coverage: what percentage of the future tumor recurrence was captured within the planned radiation volume?

  • Superiority over Standard Care: For patients with FET-PET data, GliODIL improved average recurrence coverage from 70.04% (Standard 1.5 cm margin) to 72.94%.
  • Patient-Specific Wins: In a head-to-head comparison, GliODIL provided better coverage than the standard plan for 35% of patients in the FET-PET group.
  • Outperforming Rigid Models: GliODIL consistently beat models that strictly enforced PDE constraints (like PDECMA-ES), proving that flexibility in the mathematical model is essential for clinical accuracy.
Comparison of a standard uniform margin plan vs. the GliODIL plan. GliODIL captures a significantly higher percentage of the actual tumor recurrence (74.37% vs 31.11% in this specific case).

Limitations and Future Work

  • Anatomical Barriers: The current model uses a simplified tissue map. The authors suggest that incorporating Diffusion Tensor Imaging (DTI) to account for brain fiber orientation could further refine the predicted migration pathways.
  • Initial Conditions: The model currently assumes a single focal start point. While it can handle some multi-focal complexity, a more sophisticated initial condition setup is needed for highly complex cases.
  • Single Time-Point: The study focused on immediate post-operative planning using a single pre-operative scan. Future work could leverage time-series imaging to better calibrate growth parameters.

Summary updated

Abstract

Brain tumor growth is unique to each glioma patient and extends beyond what is visible in imaging scans, infiltrating surrounding brain tissue. Understanding these hidden patient-specific progressions is essential for effective therapies. Current treatment plans for brain tumors, such as radiotherapy, typically involve delineating a uniform margin around the visible tumor on pre-treatment scans to target this invisible tumor growth. This "one size fits all" approach is derived from population studies and often fails to account for the nuances of individual patient conditions. We present the GliODIL framework, which infers the full spatial distribution of tumor cell concentration from available multi-modal imaging, leveraging a Fisher-Kolmogorov type physics model to describe tumor growth. This is achieved through the newly introduced method of Optimizing the Discrete Loss, where both data and physics-based constraints are softly assimilated into the solution. Our test dataset comprises 152 glioblastoma patients with pre-treatment imaging and post-treatment follow-ups for tumor recurrence monitoring. By blending data-driven techniques with physics-based constraints, GliODIL enhances recurrence prediction in radiotherapy planning, challenging traditional uniform margins and strict adherence to the Fisher-Kolmogorov partial differential equation model, which is adapted for complex cases.
BibTeX
@inproceedings{balcerak2025individualizing,
  author = {Michal Balcerak and Jonas Weidner and Petr Karnakov and Ivan Ezhov and Sergey Litvinov and Petros Koumoutsakos and Tamaz Amiranashvili and Ray Zirui Zhang and John S. Lowengrub and Bene Wiestler and Bjoern Menze},
  title = {{Individualizing Glioma Radiotherapy Planning by Optimization of Data and Physics-Informed Discrete Loss}},
  booktitle = {Nature Communications},
  year = {2025},
  doi = {10.1038/s41467-025-60366-4},
}
Figure