AI Models & Architecture
Our integrated framework combines three advanced AI models: segmentation, surrogate CFD, and risk prediction to deliver comprehensive cerebrovascular analysis.
Complete Analysis Pipeline
Deep learning model extracts 3D vascular geometry from MRA scans with 98% accuracy.
- Detects aneurysms and stenosis
- 3D vascular reconstruction
- Automated morphology extraction
Neural surrogate rapidly estimates pressure, velocity, and WSS in real-time.
- 1000× faster than traditional CFD
- Predicts WSS, pressure, velocity
- 98%+ accuracy vs full CFD
AI classifier combines morphology + hemodynamics to predict stroke risk with 89% accuracy.
- Integrates geometry + flow features
- Identifies high-risk zones
- Clinically actionable outputs
Architecture
Network: 3D U-Net with ResNet encoder
Input: Time-of-Flight (TOF) MRA volumes (256×256×128 voxels)
Output: Binary segmentation masks for vessels, aneurysms, and stenotic regions
Training Data: 1,200+ annotated MRA scans from AneuRisk and clinical datasets
Technical Details
Performance Metrics
- • Dice Coefficient: 0.92
- • Sensitivity: 98.3%
- • Specificity: 96.7%
- • Inference Time: 2.5 seconds
Key Features
- • Multi-scale feature extraction
- • Attention mechanisms for small vessel detection
- • Post-processing for topology refinement
- • Integrated with 3D Slicer workflow
Clinical Applications
Aneurysm Detection: Identifies saccular and fusiform aneurysms ≥3mm with 96.5% accuracy
Stenosis Quantification: Measures arterial narrowing percentage in major cerebral arteries
3D Reconstruction: Generates patient-specific surface meshes for CFD simulation
Research Foundation
Based on Bernecker et al. (2025) and Zhou et al. (2024), achieving participant-wise accuracies of 76.5% for simultaneous aneurysm and stenosis detection using patch-wise residual neural networks.
System Architecture & Integration
Infrastructure
- Framework: PyTorch 2.1 with CUDA 12.1
- Deployment: ONNX Runtime for cross-platform inference
- Processing: 3D Slicer for segmentation, SimVascular for mesh generation
- Hardware: NVIDIA A100 GPU (40GB VRAM)
Data Pipeline
- Training Datasets: AneuRisk, ADAM, MIDAS, clinical data (N=2,730)
- Preprocessing: Intensity normalization, N4 bias correction, resampling to 0.5mm³
- Augmentation: Random rotation, elastic deformation, intensity jittering
- Validation: 5-fold cross-validation with external test set
End-to-End Processing Time
2.5s
Segmentation
15s
CFD Surrogate
1.2s
Risk Prediction
~19s
Total Pipeline
Research Foundation
Ezz Zarif, N., Abdelzaher, J.G., Sakr, A.M., et al. (2025)
"AI-Enhanced Stroke Risk Prediction Using Brain MRA and Surrogate CFD Modeling"
IEEE Conference Template
Shi, Z., et al. (2021)
"Machine learning for prediction of aneurysm rupture using hemodynamic and morphological features"
Rygiel, P., et al. (2025)
"Geometric deep learning for wall shear stress prediction in cerebrovascular models"
Ready to Analyze Your MRA Scans?
Try our integrated AI pipeline with sample data or upload your own cerebrovascular MRA scans.