How arturia-pendragon-iris/hipas_av_segmentation works
Main Components Identification
a) HiPaS (Core Segmentation Module)
b) Reconstruction Module
c) Analysis Module
Detailed Architecture Breakdown
A. Data Flow Layer
• Input: CT Scans (HDF5 format)
• Data Preprocessing: Normalization (-1000, 600) to (0, 1)
• Output: Segmented Arteries and Veins
B. Processing Components
HiPaS Core (Main Module)
• Training Pipeline
• Dataset Management (datasets/)
• Data Augmentation (augment/)
• 3D UNet Architecture (unet3d/)
• Model Configuration (train_config_segmentation.yaml)
• Prediction Pipeline (predict_av.py)
Reconstruction Module
• Custom Loss Functions (losses.py)
• Reconstruction Model (model.py)
• Reconstruction Process (reconstruct.py)
Analysis Module
• Connectivity Analysis (connectivity.py)
• Distance Transform (distance_transform.py)
• Skeleton Analysis (get_skeleton.py)
• Surface Analysis (get_surface.py)
Use the following visual elements:
• Blue rectangles: Main modules
• Green rectangles: Sub-modules
• Yellow diamonds: Data processing steps
• Red hexagons: Model components
Key Relationships to Highlight
a) Training Flow:
Raw CT Data → Data Preprocessing → Training Pipeline → Model Training → Trained Model
b) Inference Flow:
Input CT → Preprocessing → HiPaS Model → Segmentation → Reconstruction → Analysis
Technical Details
• Input specifications: 512x512x512 normalized CT scans
• Model architecture: Modified 3D UNet
• Output format: Two-channel segmentation (arteries and veins)
• Processing time: ~2 minutes per CT volume
Implementation Notes
• The modular nature of the system
• Clear separation between training and inference pipelines
• Data transformation steps
• Integration points between modules