Experimental · open source

Follow a fiber.
Inspect every proposal.

Two endpoints and a small CT region produce a candidate fiber trace. This prototype tests whether image guidance can improve on a straight line—and makes its failures visible.

Measured on two additional scan regions

Parameters were frozen before evaluation on these regions. The values below measure agreement with existing human annotations, in voxels.

Region	Segments	Straight line	Proposed trace
Scroll 1, second cube	275	2.410	2.020
Scroll 5, first cube	142	3.206	2.682

Mean symmetric annotation distance decreased by approximately 16% in each cube. 112 of 417 cases worsened. The interactive review includes the largest improvement, median change, and largest regression in each region.

What this does not establish

This is a comparison with straight-line interpolation, not current Vesuvius tracing software. Annotation agreement does not prove that a path stays on the same winding. Strongly curved fibers can pull the proposal onto neighboring structures. No annotator time saving, calibrated confidence, or native VC3D import has been demonstrated.

With one extra control point

The latest diagnostic gives both methods the same middle waypoint. Away from the supplied controls, mean reference-to-path error falls from 1.416 to 1.194 voxels on Scroll 1 and from 1.769 to 1.376 on Scroll 5. The image-guided path improves in 312 of 417 cases.

Adding a waypoint to a straight line helps more than image guidance with just two endpoints. Version 1.1 therefore supports ordered waypoints in one command. These tests reuse the previous cubes and substitute existing annotation nodes for human input; they do not measure annotator time or compare against VC3D. Read all four comparisons.

Reproduce and inspect

Download the source archive. Use Python 3.14 in a virtual environment and install its pinned requirements.
Run python download_data.py to fetch approximately 53 MB of official public inputs. SHA256 hashes and file sizes are checked before use.
Run the included tests, held-out evaluation, and independent-region evaluation as described in the README.
Use the CLI with local TIFF or Zarr data to export a candidate trace as XYZ JSON and NML. Inspect raw slices before accepting any proposal.

Per-case results · Release checksums · MIT source license

Provenance

Data used in this research include EduceLab-Scrolls and Vesuvius Challenge CT scans, with archived Vesuvius fiber annotations. Source and evaluation were developed using an autonomous AI coding assistant.

Original code is MIT licensed. Cropped scan illustrations and annotations retain their separate CC-BY-NC 4.0 and dataset attribution terms. See Parsons et al., EduceLab-Scrolls (2023). The original full scans are not bundled.