TomoImageStitcher

TomoImageStitcher — Documentation

GPU-accelerated, sub-pixel accurate 3D volumetric stitcher for tomographic and large-volume microscopy datasets. Originally developed for stitching local X-ray tomography volumes for the Experiment at the DanMAX beamline, Sweden.

Start here

The best way to learn the pipeline is the detailed walkthrough notebook — it covers all six stages with explanations, parameter choices, and visualisations on a synthetic dataset:

• ../notebooks/02_full_pipeline.ipynb

The mathematical details of the registration (ZNCC, IC-GN Lucas–Kanade, mask weighting) and blending (distance-map weighting) are described in [PUBLICATION].

Contents

Page	Description
Installation	Step-by-step install: drivers, CuPy, the package itself.
Quickstart	Copy-paste recipes for the most common workflows.
Architecture	Data structures, pipeline stages, coordinate conventions.
API reference	Public classes, methods, and parameters.
Troubleshooting	Common errors and how to recover.

Quick links

• GitHub repository
• Project layout
• Citation
• Detailed walkthrough notebook

Why TomoImageStitcher?

Most off-the-shelf stitching tools (ImageJ/Fiji Grid/Collection stitching, BigStitcher, etc.) are designed for 2D tiles. TomoImageStitcher is built specifically for 3D sub-volumes with the following goals in mind:

• Sub-pixel registration via a ZNCC pixel search followed by an inverse-compositional Gauss–Newton (IC-GN) Lucas–Kanade refinement.
• GPU acceleration of every heavy step (correlate, Lucas–Kanade, Gaussian filtering, affine transform) through CuPy.
• Mask-aware interpolation so that background (zero) pixels never bleed into the foreground when blending.
• Affine or rigid transformations per pair, with optional extraction of the rigid component.
• Per-layer batching to deal with stage-z (height) stratification and rotation stages.
• Intensity equalisation through joint histograms of overlapping regions.

This site is open source. Improve this page.