Lumen

Train endovascular navigation against the vessel wall, not just the target.

Lumen is an Apache-2.0 simulation and RL environment for catheter and guidewire navigation. It combines deformable vascular anatomy, finite-radius contact, synthetic imaging, replayable episodes, and Gymnasium tasks in one GPU-parallel solver.

Launch page · Preprint · Solver coverage

Benchmark Snapshot

The historical 50,000-step branch-navigation pilot measured 100% raw target reach in both native environments and 79.7 versus 12.1 evaluation steps/s for Lumen versus CathSim. Its safety fields are not cross-environment comparable: Lumen recorded centerline penetration in simulator units while CathSim recorded native contact force. The current benchmark contract reports those endpoints separately; no cross-environment safety claim is made until matched device/anatomy/material calibration and force–injury validation exist. The frozen pilot artifacts remain available for provenance and must not be reinterpreted under the new contract.

The full preprint and benchmark summaries are linked from the launch page.

The scaled protocol is an explicit disjoint-case check:

lumen benchmark bench_results --suite scaled --episodes 100

It evaluates the same policy on frozen procedural train and held-out case IDs, reports raw/native-safe/crash rates for each split, and records the train-minus-held-out generalization gap in forward-baseline-heldout.json.

Canonical scorecards also include episode-level lumen-stats/1 summaries: fractional interquartile means, ordinary means, and deterministic 95% percentile-bootstrap intervals with recorded seeds and resample counts.

Generated scorecards include a SHA-256 certificate for every action sequence and native episode outcome. Re-run the canonical baseline certificate locally with:

python tools/check_replay.py

Use lumen.bench.replay_verified_leaderboard(results_dir, policies) to rank only scorecards whose certificates re-run successfully with the supplied policy callables.

Install

git clone https://github.com/SeldingerMed/seldinger-lumen
cd seldinger-lumen
pip install -e ".[dev]"
lumen doctor

CPU Docker image

The repository ships a CPU-only runtime image. It installs the pinned Newton/Warp solver but does not require a CUDA device:

docker build --file docker/Dockerfile --tag seldinger-lumen:0.2.0 .
docker run --rm seldinger-lumen:0.2.0 anatomy --validate

First Run

lumen play stenotic --out lumen-run
lumen benchmark lumen-bench
lumen anatomy --validate
lumen render-fluoro lumen-fluoro.png
lumen capture lumen-episodes
lumen validate lumen-episodes --require-cv-labels
lumen index lumen-episodes --out lumen-episodes/index.jsonl --check-sidecars
lumen split-index lumen-episodes/index.jsonl --out-dir lumen-episodes/splits

Python API

import gymnasium as gym
import lumen.envs.registration

env = gym.make("Lumen/NavStenotic-v0")
obs, info = env.reset()
obs, reward, terminated, truncated, info = env.step(env.action_space.sample())

For standard RL libraries, the optional adapters preserve the Gymnasium contract:

from lumen.rl import make_cleanrl_vector_env, make_sb3_env

sb3_env = make_sb3_env("Lumen/NavStenotic-v0", seed=0)
cleanrl_envs = make_cleanrl_vector_env("Lumen/NavStenotic-v0", num_envs=4, seed=0)

Install Stable-Baselines3 or CleanRL separately when using those trainers; the core package keeps both integrations optional.

Core systems

• Procedural stenotic, tortuous, aneurysmal, and branching vessels.
• Fixed-port guidewire and coaxial catheter actuation with rotation, latency, backlash, and motion limits.
• Deformable wall mechanics, finite-radius contact, anisotropic friction, flow, clot, retrieval, and flow diversion.
• Synthetic fluoroscopy, luminal RGB, masks, keypoints, noise, latency, and dropout.
• Dataset capture, validation, replay, indexing, splitting, and materialization.
• Privileged, tracked, and raw-image observation contracts for RL.
• Simulator/phantom/device deployment interface with fail-closed safety envelopes and force/torque benchtop validation (protocol).

Citation

@software{son_lumen_2026,
  author = {Son, Colin},
  title = {Lumen: an Open, Differentiable, GPU-Parallel Environment for Endovascular AI},
  year = {2026},
  url = {https://github.com/SeldingerMed/seldinger-lumen},
  license = {Apache-2.0}
}

License

Apache-2.0.