Industrial CT volumes, sieved.

VoxelSieve removes the outside air from reconstructed CT scans, keeps the part and its internal voids, and stores the result as a sparse OpenVDB dataset. On top of it: porosity analysis, inspection reports after BDG P 202, nominal-actual comparison with the CAD model, dimensional measurement with first article reports, a browser UI and an MCP server for AI assistants. Scans larger than memory work, streamed out of core.

Quickstart Download Source on GitHub

Apache-2.0 · Linux, Windows, macOS · C++20

Deviation of a warped housing with a dent from its CAD model, coloured by deviation

What it does

Sieve

Raw volumes (with vendor headers) and TIFF stacks, also inside ZIP archives, become a bricked VDB dataset. Bricks with only outside air are not written. Grey values stay lossless.

Porosity

Finds pores and zones of loosened microstructure, with volumes from the grey values and cupping kept apart from porosity. On a synthetic scan of a cast bell housing every shrinkage cavity is within 2 % of the ground truth.

Inspection report

Evaluates the analysis against acceptance limits per inspection zone (BDG P 202 scheme) and writes an HTML report to print as PDF.

Dimensional measurement

Measures the characteristics of a drawing on the scan: planes, cylinders and spheres found on the CAD model, datums 3-2-1 or RPS, sizes, form, orientation, position and profile against their tolerances.

First article report

Cover sheet, characteristic list with nominal, tolerance, actual value and OK/NOK, and views of the part with the balloon numbers, made by an AI assistant from scan and CAD model.

Denoising

Non-local means against the measured noise of the scan, or a learned model trained on the noisy scan alone, into a new dataset that keeps the part.

Nominal-actual

Aligns the CAD model (STL) to the scanned surface, from any pose, and measures the signed deviation of every surface point, with statistics and a coloured mesh.

GUI

A browser UI with slice and 3D views that load only what is on screen, a wizard, undo and redo, saved views and several objects in one coordinate system.

Learned segmentation

Material classes by grey value or a learned model run on the CPU tile by tile, scored against labels. Training scripts in PyTorch are included.

Bell housing projected along its axis, pores in red and loosened zones in yellow — Porosity: pores in red, loosened zones in yellow.

Evaluation table of the inspection report with three inspection zones — Inspection report: evaluation per zone.

First article inspection report of a wheel hub: cover sheet, characteristic list with one characteristic out of tolerance, and two views of the part with balloon numbers — First article report: characteristics measured on the scan, out of tolerance in red.

3D view of a cast bell housing with pores and loosened zones inside the translucent volume — 3D view: pores and loosened zones inside the translucent part.

GUI in the browser with the step protocol and a porosity result — The GUI: every step in the protocol, with undo and redo.

Slice view zoomed onto a shrinkage cavity next to a loosened zone in a flange — Slice view, zoomed onto a shrinkage cavity and a loosened zone.

Works with AI assistants

vs-gui --mcp serves VoxelSieve over the Model Context Protocol. An assistant such as Claude Desktop or Claude Code can create projects, run the analyses, look at slices and write the report, up to a first article inspection report from a scan and its CAD model.

It runs on your computer: the scans stay where they are, and the assistant only receives results such as pore lists, measurements, slice pictures and reports.

claude mcp add voxelsieve -- \
  /opt/voxelsieve/bin/vs-gui --mcp /data/scans

> Import /data/scans/housing.raw, analyse the
  porosity and write an inspection report with
  the example order.

Get started

Download a package from the latest release: an MCP bundle (.mcpb) for Claude Desktop, or an archive with all tools for Linux, Windows and macOS. The Docker image is ghcr.io/yoshi5534/voxelsieve.
Get the sample data: a synthetic scan of a cast housing with its CAD model, an inspection order and the ground truth of its defects.
Follow the quickstart from download to a porosity report and a CAD comparison in about ten minutes.

To build from source, see the README; dependencies come from vcpkg. Contributions are welcome, see CONTRIBUTING.md.