Federated PINNs achieve 91.4% accuracy in brain tumor modeling without sharing patient data
New method combines physics and privacy for brain tumor biomechanics with near-perfect AUC.
Deep Dive
A federated physics-informed neural network (PINN) was introduced for biomechanical modeling of brain tumors. Training across three simulated clinical sites with only model weights shared via FedAvg achieved 91.4% accuracy versus 90.0% for a non-federated baseline trained on pooled data, an average AUC of 0.985 across tumor classes, and a rise in pituitary tumor accuracy from 85.6% to 94.5% while keeping patient MRI data local, addressing privacy rules such as GDPR and HIPAA.
Key Points
- Federated PINN combines federated learning with linear elasticity physics-informed loss to model brain tumor biomechanics from MRI data
- Achieved 91.4% accuracy and 0.985 average AUC, with pituitary tumor accuracy rising 8.9% over non-federated baseline
- Only model weights shared via FedAvg protocol across 100 rounds; raw patient data stays local, fully HIPAA/GDPR compliant
Why It Matters
Enables hospitals to collaborate on brain tumor AI models without sharing sensitive patient data, improving accuracy and generalization.