Seldinger NeuroCathSim Policy
This repository contains a Seldinger NeuroCathSim command-conditioned policy checkpoint for anterior-circulation simulated endovascular navigation through the repository's CathSim navigation sweep harness.
The released checkpoint is image-conditioned: at inference it consumes the grounded command plus DSA-like rendered image observations rather than graph-only simulator state features.
It is a non-clinical research artifact. It must not be used for patient care, autonomous clinical navigation, procedural advice, or medical device selection.
Artifact
- Hugging Face repo:
txmedai/seldinger-neurocathsim-policy
- Checkpoint:
rl_policy.json
- Source commit:
d333dff717ced6d1912adb6b33021f60d58b2929
- Policy model type:
command_conditioned_image_linear_bc
- Observation mode:
image
- Feature mode:
image
- Feature count: 27
- Training source:
oracle_rollouts
- Target count: 12
- Success rate: 1.000
- Mean path efficiency: 1.000
- Wrong-branch count: 0.0
- Wall-contact proxy count: 0.0
- Failure cases: 0
Intended Use
The policy is intended for reproducible research in the open NeuroCathSim/CathSim integration path. It maps a grounded anterior-circulation command and simulator observation to support-catheter, microcatheter, and microwire/guidewire actions, then is evaluated by commanding the simulator to each anterior-circulation target in closed loop.
Limitations
The current release is a command-conditioned research policy artifact, not a validated clinical simulator or robot controller. The CathSim dependency is the public GitHub project at https://github.com/airvlab/cathsim; this repository adapts it through an anterior-circulation navigation sweep harness while richer neurovascular phantom physics, anatomy diversity, device-specific mechanics, and external validation remain active work.
Citation
See the Seldinger repository manuscript package for the evolving preprint source.
