VeSeg

VeSeg is a Python package for vascular segmentation and vessel geometry extraction. The package combines a pretrained U-Net with a Rust backend to generate vessel masks, skeletons, vascular graphs, and radius estimates from microscopy and visible-light vascular images.

The long-term purpose of VeSeg is to convert raw vascular images into representations usable for:

• Oxygen and nutrient transport simulations
• Tissue diffusion modeling
• Vascular topology extraction
• Vessel skeletonization and graph generation
• Vessel radius and diameter estimation
• Biomedical image analysis
• Future blood flow and coupled transport simulations

---

Pipeline

Image
↓
Preprocessing
↓
U-Net segmentation
↓
Binary vessel mask
↓
Skeletonization (Rust)
↓
Distance transform
↓
Graph extraction
↓
Radius estimation
↓
Simulation-ready geometry

---

Features

Current:

• U-Net based vessel segmentation

• Multiple preprocessing modes

  • Grayscale
  • Inverted
  • Enhanced + inverted
  • Red channel extraction

• Automatic postprocessing

• Speck removal

• Binary mask generation

• Mask scaling utilities

• Installable Python package

• Public prediction API

• Dataset augmentation pipeline

• Threshold sweep optimization during validation

• Rust-accelerated skeletonization

• Vessel graph extraction

• Vessel radius estimation

• Bundled geometry extraction backend

• Distance transform generation

• Radius-annotated edge tables

• Vessel reconstruction from skeleton + radius estimates

---

Installation

Install from GitHub

pip install git+https://github.com/HARDIntegral/VeSeg.git

Note: Prebuilt wheels and bundled binaries are currently tested primarily on Apple Silicon (ARM64 macOS). Installation on other systems may require compiling the Rust extension locally.

For non-Apple Silicon systems or development workflows, the recommended installation method is the following development installation.

Development installation

git clone https://github.com/HARDIntegral/VeSeg.git
cd VeSeg

python -m venv .venv
source .venv/bin/activate

pip install maturin
maturin develop

Build wheel:

maturin build --release --out target/wheels

Run tests:

pytest

---

Quick Start

Predict vessels and extract geometry:

import veseg

result = veseg.predict(
	"test_images/structure1.png",
	mode=veseg.ENHANCED_INVERTED,
)

mask, skeleton, distance_map, nodes, edges = (
	veseg.extract_vessel_geometry(result.mask)
)

print("nodes:", nodes.shape)
print("edges:", edges.shape)
print("mean radii:", edges[:, 4])

Scale masks:

scaled = result.scaled(2.0)
scaled.save_mask("large_mask.png")

Remove isolated false positives:

clean = result.despeckle(min_neighbors=3)

Radius estimates are returned inside the edge table:

mean_radius = edges[:, 4]
min_radius = edges[:, 5]
max_radius = edges[:, 6]

---

Public API

Available:

veseg.predict()
veseg.load_model()

veseg.extract_vessel_geometry()
veseg.build_vessel_geometry()
veseg.build_vessel_graph()

veseg.skeletonize_mask()
veseg.distance_transform_mask()
veseg.reconstruct_vessel_mask()

veseg.postprocess_prediction()
veseg.scale_mask()
veseg.resize_mask()
veseg.remove_specks()

Prediction modes:

veseg.GRAYSCALE
veseg.INVERTED
veseg.ENHANCED_INVERTED
veseg.RED

---

Model

Current architecture:

Input
 ↓
Encoder
 ↓
Bottleneck
 ↓
Decoder + skip connections
 ↓
Pixelwise logits

Model:

U-Net

Training:

Loss:
0.5 BCEWithLogits + 0.5 Dice

Optimizer:
AdamW

---

Tests

Run:

pytest

Current tests include:

• dataset loading
• metrics
• package API validation

---

Documentation

Detailed API documentation, parameters, return values, graph schemas, and examples:

DOCS.md

---

License

MIT License