SeqSeg: Sequential Vessel Segmentation and Tracking

Automatic tracking and segmentation of blood vessels in CT and MR images using deep learning and geometric tracking.

News: SeqSeg now outputs a full SimVascular project in the simvascular/ subdirectory — open it directly in SimVascular with automatic pathlines and contours for every segmented branch.

Why SeqSeg?

SeqSeg segments vessels sequentially, taking steps along vessel centerlines and detecting bifurcations to grow complete vascular trees from just 1–2 seed points. By combining local deep-learning predictions (nnU-Net) with geometric tracking, it stays robust across vessel scales — from small coronaries to large aortas.

• 🌱 Minimal supervision — only 1–2 seed points to initialize
• 🌿 Robust bifurcation detection — automatically follows every branch
• 🩻 Multi-modal — works with CT and MR 3D medical images
• 📏 Scalable — vessels from ~1mm coronaries to ~30mm aortas
• ✅ Clinically validated — coronary, aortic, cerebral, and pulmonary anatomies (pre-trained weights for aorta CT/MR and coronary CT)
• ⚡ Fast — ~2–10 min per case, Dice > 0.9 on validation, runs on CPU or GPU

Real-time demonstration: automatic segmentation of an abdominal aorta in a 3D MR scan.

What's new in 2.x

SeqSeg 2.0 refactors the package around a clearer CLI and a stable Python API. Existing batch workflows still work; legacy invocations without a subcommand (e.g. seqseg -data_dir ...) are rewritten to seqseg run batch automatically.

Command-line interface

Python library API

Embed tracing in other Python code without writing SeqSeg output files:

• seqseg.api.run_tracing — pass a sitk.Image, seed definitions, and an nnU-Net trainer folder; get a TracingResult with global probability segmentation at result.assembly.assembly
• TracingOptions(disk_io=False) — skip VTK/MHA debug trees on disk (nnU-Net weights still load from model_folder)
• BranchSeed, branch_seed_at_point, seeds_to_potential_branches — simple seed formats instead of hand-built step dicts
• TracingContext / trace_centerline_from_context — lower-level control with the same in-memory image support
• Lazy re-exports from import seqseg (see seqseg/__init__.py)

Quick example (seeds and config known):

from seqseg.api import TracingOptions, branch_seed_at_point, run_tracing

result = run_tracing(
    my_sitk_image,
    [branch_seed_at_point([x, y, z], radius)],
    "/path/to/nnUNetTrainer__nnUNetPlans__3d_fullres",
    config="global",
    options=TracingOptions(disk_io=False),
)
prob_seg = result.assembly.assembly  # sitk.Image; threshold for binary masks

See High-level API (seqseg.api) for full detail.

Internal structure (for contributors)

• Pipeline modules: seqseg.pipeline.classic, plus, post, single_trace
• Typed config helpers: AlgorithmConfig, NnUNetModelSpec in seqseg.config_models
• Tracing core accepts sitk.Image or file paths for the reference volume and optional prior segmentation

Migrating from 1.x

1. CLI: Prefer seqseg run batch (or keep legacy flags — they still work).
2. Plus workflow: Use seqseg run plus batch instead of python -m seqseg.seqseg_plus with the same nnU-Net path flags.
3. Library: Use run_tracing or TracingContext rather than calling trace_centerline with only file paths.
4. Version: pip install -U seqseg and check with seqseg --version (expects 2.1.0).

Quick Start

# Install
pip install seqseg

# Download pre-trained weights (see Installation docs for links)
# Run segmentation
seqseg -data_dir your_data/ -nnunet_results_path path/to/weights/ -config_name aorta_tutorial

📖 New here? Follow the step-by-step tutorial with example data and detailed instructions.

Documentation

Citation

When using SeqSeg, please cite the following paper:

@Article{SveinssonCepero2024,
author={Sveinsson Cepero, Numi
and Shadden, Shawn C.},
title={SeqSeg: Learning Local Segments for Automatic Vascular Model Construction},
journal={Annals of Biomedical Engineering},
year={2024},
month={Sep},
day={18},
issn={1573-9686},
doi={10.1007/s10439-024-03611-z},
url={https://doi.org/10.1007/s10439-024-03611-z},
}