Physics-informed DynUNet for brain metastasis segmentation
Physics-informed AI modestly improves brain metastasis mapping on imaging, with potential to guide surgery planning.
Integrating physics-based tumor growth modeling into a deep learning segmentation architecture (DynUNet) produced consistent but modest gains in brain metastasis segmentation across all tumor regions and a clinically meaningful +10.1% tumor-core Dice improvement in external validation (Stanford, n=105). Optimal regularization weights were context-dependent, suggesting configuration tuning would be required for deployment.
What the study was
- Study design
- Retrospective ML model development with external validation
- Population
- Brain metastasis patients (BraTS-METS 2023 + Stanford BrainMetShare)
- Sample size
- 105
- Category
- Diagnostics
- Maturity
- Exploratory
- Journal
- Computer Methods and Programs in Biomedicine
Why it surfaced
Technically sound physics-informed approach with external validation, modest gains. Comput Methods Programs Biomed.
A plain-language summary of published research — not medical advice. Talk to a clinician about your own care.