Modeling Aneurysm Hemodynamics

Typical pipeline from patient scan to risk prediction:

Python
# 1. Extract geometry from segmented vessel (DICOM -> STL)
# 2. Generate CFD mesh and run reference simulation (ground truth)
# 3. Extract geometric + hemodynamic features
# 4. Train ML surrogate to predict hemodynamic fields/metrics without full CFD

import numpy as np
from sklearn.ensemble import GradientBoostingRegressor

features = [
    'aspect_ratio', 'size_ratio', 'neck_width', 'inflow_angle',
    'wall_shear_stress_mean', 'oscillatory_shear_index', 'curvature'
]

X = df[features].values
y = df['rupture_risk_label'].values  # or continuous WSS/OSI target

model = GradientBoostingRegressor(n_estimators=300, max_depth=4, learning_rate=0.05)
model.fit(X_train, y_train)

No preamble needed — state the clinical/engineering question first (rupture risk classification vs. flow field regression), since it determines architecture choice.

Progress:

• Step 1: Define the modeling target (rupture risk classification, WSS/OSI field regression, or flow surrogate for real-time use)
• Step 2: Acquire/segment vascular geometry (CTA/MRA → 3D mesh via Vascular Modeling Toolkit or similar)
• Step 3: Generate ground-truth hemodynamics via CFD (Navier-Stokes, Newtonian or Carreau-Yasuda blood rheology, pulsatile boundary conditions from patient-specific or literature flow waveforms)
• Step 4: Extract feature set — geometric (aspect ratio, size ratio, undulation index, neck width, curvature, bifurcation angle) and hemodynamic (WSS, OSI, RRT, inflow concentration, kinetic energy)
• Step 5: Choose model architecture matched to the target:
  • Tabular risk classification → gradient boosting (XGBoost/LightGBM), random forest
  • Spatial field prediction → graph neural networks (GNN) on mesh, or PointNet/mesh-CNN
  • Full flow-field surrogate → physics-informed neural networks (PINNs) or neural operators (DeepONet, FNO)
  • Temporal/pulsatile dynamics → recurrent or temporal graph models
• Step 6: Validate against CFD ground truth — field-wise (point-wise error, relative L2 norm) and clinically relevant metrics (max WSS location, OSI hot-spot overlap)
• Step 7: Cross-validate on patient cohort with leave-one-patient-out, not random split (avoid data leakage between same-patient samples)
• Step 8: Quantify uncertainty (ensemble variance, MC dropout, conformal prediction) — critical for clinical decision support
• Step 9: Report with physical plausibility checks (mass conservation, non-negative pressure where expected)

Example 1: Input: Dataset of 150 patient-specific cerebral aneurysms with CFD-computed WSS fields, geometric parameters, and clinical rupture status (ruptured/unruptured). Output: A pipeline using random forest classifier on 12 geometric + hemodynamic features achieving AUC 0.84 (leave-one-patient-out CV), with SHAP analysis identifying low WSS + high OSI + large aspect ratio as top rupture predictors; recommendation to validate feature importance against published PHASES/ELAPSS clinical scores.

Example 2: Input: Need for real-time (sub-second) hemodynamic field estimation for intraoperative decision support, given only vessel geometry mesh. Output: Graph neural network (GNN) trained on 200 CFD simulations as surrogate; input = mesh node coordinates + inlet flow rate, output = per-node WSS vector; architecture = 6-layer message-passing GNN with edge features from mesh connectivity; achieves relative L2 error <8% vs. CFD at ~50ms inference vs. 4hr CFD; flag that extrapolation to unseen vessel topologies requires retraining or domain adaptation.

Example 3: Input: Small dataset (30 patients) — insufficient for pure deep learning. Output: Reduced-order physics-informed model: fit low-dimensional Navier-Stokes approximation (e.g., 1D reduced-order hemodynamic model or POD-based reduced basis) with ML-tuned closure parameters instead of end-to-end black-box; preserves physical interpretability and generalizes better with limited data.