Model Workflow

From MRA scan to stroke risk prediction in under 20 seconds. See how our three AI models work together seamlessly.

Input: MRA Scan Upload

0s

Clinician uploads Time-of-Flight (TOF) MRA scan in DICOM (.dcm) or NIfTI (.nii) format.

Format Support

.dcm, .nii, .nii.gz

Resolution

0.5mm isotropic voxels

Typical Size

40-60 MB per scan

Model 1: Vessel Segmentation

2.5s

3D U-Net with ResNet encoder extracts cerebrovascular anatomy from MRA volume.

Processing Steps

  1. 1.MRA volume normalized and preprocessed (intensity scaling, bias correction)
  2. 2.3D U-Net performs voxel-wise classification (vessel vs. background)
  3. 3.Post-processing removes noise and refines topology
  4. 4.Binary mask converted to 3D surface mesh for visualization

Output

  • • Binary segmentation mask
  • • 3D vascular surface mesh
  • • Aneurysm location coordinates
  • • Stenosis regions identified

Key Metrics

  • • Dice Score: 0.92
  • • Sensitivity: 98.3%
  • • Specificity: 96.7%
  • • False Positive Rate: <3%

Model 2: Surrogate CFD

15s

Physics-informed neural network computes hemodynamic biomarkers 1000× faster than traditional CFD.

Hemodynamic Simulation

  1. 1.Segmented mesh converted to computational domain with boundary conditions
  2. 2.PINN predicts velocity field (u, v, w components) at each mesh node
  3. 3.Pressure distribution calculated from Navier-Stokes momentum balance
  4. 4.Wall Shear Stress (WSS) computed as tangential velocity gradient at vessel walls

Velocity Field

Blood flow speed (cm/s)

Normal: 20-40

Stenosis: >80

Pressure

Blood pressure (mmHg)

Systolic: 120

Gradient: 10-15

WSS

Wall shear stress (Pa)

Normal: 1-7

Risk: <0.4 or >10

Model 3: Risk Prediction

1.5s

Ensemble classifier combines morphological + hemodynamic features to predict stroke risk.

Feature Integration & Classification

  1. 1.Extract 47 morphological features (aneurysm size, neck width, vessel tortuosity)
  2. 2.Extract 38 hemodynamic features (max/min WSS, pressure gradient, OSI)
  3. 3.Ensemble of XGBoost + Random Forest classifiers predicts rupture probability
  4. 4.Attention mechanism highlights high-risk vascular regions

Risk Categories

  • • Low: <20% rupture risk
  • • Moderate: 20-50% risk
  • • High: 50-75% risk
  • • Critical: >75% risk

Performance

  • • Accuracy: 89.3%
  • • AUC-ROC: 0.94
  • • Sensitivity: 91.5%
  • • Specificity: 87.2%

Output: Clinical Report

Total: ~19s

Comprehensive stroke risk assessment with interactive 3D visualization and clinical recommendations.

Visualization Dashboard

  • 3D vascular reconstruction with full rotation/zoom
  • Color-coded hemodynamic maps (velocity, pressure, WSS)
  • Highlighted high-risk regions with annotations
  • Pulsatile flow animations and streamlines

Clinical Insights

  • Overall stroke risk score (0-100)
  • Quantitative hemodynamic biomarkers
  • Identified pathologies (aneurysm, stenosis)
  • Evidence-based clinical recommendations

~19s

Total Processing Time

3

AI Models in Pipeline

89%

Risk Prediction Accuracy

1000×

Faster than Traditional CFD