Written and maintained by CASRAI Editorial Board
Last updated
Image stitching is the computational joining of overlapping image tiles, for example those captured by a motorized microscope stage, into one larger image. Software estimates each tile’s position from the image content because stage coordinates are not precise enough, then blends the overlaps into a single mosaic.
Short answer: what is image stitching?
In microscopy, image stitching (also called tile stitching or mosaic assembly) takes a set of overlapping fields of view and produces one seamless image covering an area larger than a single field. The Fiji Stitching plugins describe it as combining overlapping tiles into a single 2D to 5D image. The same word is used for panorama photography; this guide covers only the microscopy and bench-imaging sense.
Definition and scope
A scan produces many tiles with a deliberate overlap. The Preibisch, Saalfeld and Tomancak paper that introduced the Fiji stitching plugins states the core problem plainly: the physical coordinates provided by the microscope stage are not precise enough, so tile positions must be refined from the pixel data themselves. Stitching is therefore a registration problem followed by an image-compositing step, not simple concatenation of files in stage order.
Stitching is separate from opening vendor file formats, which is covered in opening proprietary microscope files, and separate from the downstream measurement. It is part of the image-data record that a methods section should be able to reproduce; see the dictionary entries on metadata and reproducibility.
How image stitching is used in practice
Both the Fiji documentation and the Bioinformatics paper describe the same stages.
1. Pairwise registration
For each pair of neighbouring tiles, the software finds the shift that best aligns their overlap. Fiji uses Fourier shift (phase) correlation, evaluates several candidate peaks, and can refine the result to subpixel accuracy.
2. Global optimization
Instead of placing tiles one after another, which compounds small errors, the plugin solves a global least-squares problem over all pairwise offsets. Links that disagree strongly with the rest are removed and the solution is recomputed. Mohammadi et al. (2024) describe the same two-step structure and note that pairwise registration carries most of the computation.
3. Fusion and blending
Overlapping pixels are combined. Fiji offers linear blending, average, median, maximum and minimum fusion. In linear blending, the weight of each tile falls off from its edge; the blending fraction defaults to 20% of the tile edge. The 2009 paper explains that tunable weighting is there to handle brightness differences caused by shading.
Types of registration approach
The 2024 Scientific Reports study compared region-based and feature-based pairwise registration on three datasets (bright-field, colon fluorescence and stem cell colonies). All methods in the study assumed translation-only motion.
| Approach | What it does | Reported finding (Mohammadi et al. 2024) |
|---|---|---|
| Normalized cross-correlation (NCC) | Region-based: compares intensity patterns in the overlap | On shaded (uncorrected) tiles, 10.48% of translations were invalid |
| Phase-NCC (phase correlation) | Region-based: Fourier-domain correlation, the family used by Fiji’s pairwise step | On shaded tiles, 88.02% of translations were invalid; on shading-corrected tiles, 6.18% |
| Feature-based (Harris, FAST, ORB, BRISK, SURF, SIFT, KAZE, MSER, SuperPoint) | Detects and matches keypoints in the overlap | Described as highly robust to uneven illumination; SURF was most effective; FAST was fastest but failed on 19.28% of pairs |
The authors caution that results come from three datasets and that their MATLAB timings depend on the computing environment, so these figures show a pattern rather than a ranking that transfers to every sample.
Where it matters for labs: an acquisition and reporting checklist
The items below are drawn from the Fiji plugin parameters and the two papers; each ties a bench decision to a documented stitching behaviour.
| Decision | Why it matters (source) |
|---|---|
| Plan and record tile overlap | Fiji asks for an overlap estimate. Overlaps in the 2024 datasets ranged from about 25% (bright-field) to 2-3% (colon fluorescence) and 10-19% (stem cell colonies), so overlap is a setting to record, not a constant. |
| Correct shading before stitching | The 2024 study found phase-based registration failed on most uncorrected shaded tiles but mostly succeeded after correction. |
| Keep layout metadata (grid, filenames, tile configuration) | Fiji’s Grid/Collection Stitching works best with an approximately known layout; with no known layout all pairs must be tested and runtime grows (a 63-tile 3D set took about 3 hours in the 2009 paper). |
| Note stage axis conventions | Fiji exposes options to invert X/Y and to ignore the Z stage position, so mismatched conventions are a known setup issue. |
| Check the regression threshold | Fiji reports a regression value per link; values above about 0.7 typically indicate good registration. |
| Choose memory options deliberately | Fiji offers downsampling, virtual stacks and writing to disk as memory trade-offs for large mosaics. |
| Report software and parameters | Because positions are estimated from content, the software, version and settings (peaks, subpixel, blending) determine the mosaic. Record them with the image metadata. |
Common misunderstandings
- “The stage coordinates are enough.” The 2009 paper states they are not precise enough; content-based refinement is the point of stitching.
- “Blending removes the problem.” Blending smooths brightness differences at seams. It does not correct misregistration or uneven illumination in the underlying tiles, which is why shading correction comes first.
- “It handles any motion.” The Fiji plugins and the compared methods use a translation-only model, suited to fixed samples on level slides.
- “Every tile will be placed.” Tiles with no content or insufficient overlap are disconnected from the others and excluded.
FAQ
What is image stitching?
It is the joining of overlapping image tiles into one larger image. In microscopy, software registers neighbouring tiles by their image content, solves for consistent global positions, and blends the overlaps.
Why not just use the stage positions?
Stage coordinates are not precise enough to place tiles at pixel accuracy, so overlap content is used to refine positions.
How much overlap is needed?
The sources give no single value. Datasets in the 2024 study ranged from 2-3% to about 25%; the Fiji plugin takes your overlap estimate as an input.
Is the Fiji stitching tool free?
The Fiji Stitching plugins are open source and ship with Fiji.
Sources
- ImageJ wiki, Image Stitching (Fiji Stitching plugins): imagej.net/plugins/image-stitching, read 2026-11-08.
- Preibisch S, Saalfeld S, Tomancak P. Globally optimal stitching of tiled 3D microscopic image acquisitions. Bioinformatics 2009;25(11):1463-1465. doi:10.1093/bioinformatics/btp184. PMC2682522, read 2026-11-08.
- Mohammadi et al. Scientific Reports 2024;14:9215. doi:10.1038/s41598-024-59626-y. PMC11035624, read 2026-11-08.
