Reviewers, editors, and lab members increasingly need to look at a submitted figure and ask a specific question: is this image processed in a way that is normal and disclosed, or has it been altered in a way that misrepresents the underlying data? This guide is a practical procedure for that question — what to look for, how to check it without specialist software, and what the finding means once you have it. It covers photographic images (Western blots, gels, fluorescence and light micrographs) — the category almost all image-integrity screening and guidance targets. It does not cover line art, charts, or graphs, which raise different (usually statistical, not forensic) integrity questions.
Quick-reference: manipulation categories and how to check each one
| Category | What it looks like | How to check it | Typical editorial disposition |
|---|---|---|---|
| Internal duplication | The same band, cell cluster, or field reused within one figure or across figures — sometimes after rotation, cropping, flipping, or a contrast shift | Flip/rotate/invert suspect regions and overlay them (e.g. Photoshop’s Difference blend mode, or ImageJ/Fiji image-calculator subtraction) to test for an exact or near-exact match | Author correction if unintentional and it doesn’t affect conclusions; expression of concern or retraction if it does, or if intent is in question |
| External duplication | A submitted image matches one already published elsewhere, including in the same authors’ other papers | Reverse image search and literature-database checks; this is the category automated screening tools are strongest at, since it requires a large reference index a human reviewer doesn’t have | Depends on whether it’s disclosed reuse (e.g. a shared control image, properly cited) or undisclosed reuse presented as new data |
| Splicing / composite assembly | A figure assembled from multiple source images presented as one continuous panel, without a dividing line or disclosure | Look for straight-line borders, abrupt changes in background texture or grain/noise pattern, or misaligned edges at a panel boundary; adjusting brightness/contrast can make a disguised seam visible | Journals generally require visible dividing lines and disclosure in the legend for genuinely composite panels (e.g. non-adjacent gel lanes) — undisclosed splicing is treated as the manipulation itself, independent of whether the underlying data was accurate |
| Selective or excessive enhancement | Brightness, contrast, or color adjustment applied to only part of an image, or pushed far enough to add or remove a feature (e.g. erasing a faint background band) | Compare the processed image against the original/raw file if available; check whether an adjustment appears applied uniformly across the whole image or only to a sub-region | Correction if the effect on the paper’s conclusions is minor; more serious action if the adjustment changed what the data appears to show |
The governing standard: adjustments must apply to the whole image, not selectively
The reference point nearly every current publisher policy still traces back to is a 2004 Journal of Cell Biology editorial by Mike Rossner and Kristina Yamada, “What’s in a picture? The temptation of image manipulation” (J Cell Biol 166(1):11-15). Its core rule is simple to apply: any brightness, contrast, or color adjustment is acceptable only if it is applied to every pixel in the image, equally — not to a selected region. The moment an adjustment is applied selectively (brightening one lane, darkening a background patch, erasing a smudge), it stops being routine image processing and becomes a data-altering step that needs to be disclosed, justified, or removed.
ICMJE’s research-misconduct guidance treats this the same way: its definition of falsification explicitly includes “deceptive manipulation of images” alongside fabricating or altering data, not as a separate, lesser category. In the US federal definition that governs PHS-funded research (42 CFR Part 93), falsification is defined as manipulating research materials, equipment, or processes, or changing or omitting data or results, such that the research is not accurately represented in the research record — a definition broad enough to cover image manipulation without needing a figure-specific clause.
Step-by-step: inspecting a figure yourself
- Get the original file if you can. A screenshot or a figure cropped into a manuscript PDF has already lost information a forensic check needs (resolution, metadata, sometimes color depth). Many journals now require raw, uncropped images at submission or revision specifically so this check is possible — PLOS, for example, requires a compiled PDF of original blot/gel images with loading order and sample identity annotated for manuscripts submitted from April 2026 onward.
- Scan for duplication first — it’s the most common flag. Look across all panels in the figure, and across other figures in the same manuscript, for a repeated element: a blot band, a lane, a field of cells, a section of tissue. Duplication survives rotation, flipping, and cropping, so don’t rule out a match just because two regions look mirrored or resized relative to each other.
- Test a suspected match rather than eyeballing it. Crop both regions, align them, and overlay them — Photoshop’s Difference blend mode (an exact match produces a solid black result) or ImageJ/Fiji’s image-calculator subtraction function both do this without specialist forensic software. Free browser-based tools (e.g. Forensically) offer similar overlay and error-level-analysis functions for a quick first pass.
- Check panel boundaries for splicing. Zoom in on any line dividing sub-images within a single panel. A genuine, disclosed composite (common with gels, where non-adjacent lanes are run on the same gel and reassembled) should have a visible dividing line and a note in the legend. An undisclosed splice often shows a subtle mismatch in background texture, grain, or lighting direction on either side of the seam.
- Check contrast and brightness for selective application. Adjust the image’s own brightness/contrast sliders, or examine its histogram, looking for a hard edge where the tonal range changes abruptly within one continuous background — a sign that an adjustment was applied to part of the image rather than the whole thing.
- Weigh what you found against whether it affects the paper’s conclusions. This is the question COPE’s own decision tool is built around (see below) — not every flagged image is misconduct, and not every genuine duplication changes what the data shows.
Manual tools vs. AI-assisted screening software
The techniques above are things a reviewer, editor, or lab member can do with software most institutions already have (ImageJ/Fiji is free; Photoshop or GIMP cover the rest). They do not scale to submission volume, which is why a growing number of publishers now run figures through automated pattern-matching software before or during peer review. Two vendors — Proofig AI and ImageTwin — dominate current publisher adoption, and recent, named, publicly documented deployments give a concrete sense of what these tools actually check and catch:
- MDPI signed a multi-year agreement with Proofig AI in September 2025 to screen submissions across its biomedical journals.
- The American Society for Microbiology (ASM) published a peer-reviewed account, in mBio, of a one-year ImageTwin pilot: 2,627 accepted, eligible manuscripts were screened; ImageTwin flagged image-duplication concerns in 3.9% of them; most were unintentional and resolved through author correction; six manuscripts (0.23% of those screened) had their acceptance revoked because the issue couldn’t be resolved. The tool screens halftone images (microscopy, gel, immunofluorescence) specifically and does not evaluate line art or graphs.
Both vendors’ published capability lists group detection into the same four categories: internal duplication (including after rotation/cropping/contrast changes), external duplication against indexed literature (including PubMed Central), manipulation/splicing signatures, and — the least mature category industry-wide — undisclosed AI-generated content, where a diffusion-model-generated blot has no duplicate source image to match against in the first place. See CASRAI’s reporting on AI image-integrity screening going mainstream at MDPI and ASM for the full detail behind both figures above, and the synthetic image and AI image generation dictionary entries for the underlying generative-AI detection problem.
What to raise with an editor, and how it gets decided
Flagging a concern — to a corresponding author, an editor, or an institutional research integrity office — is a separate step from deciding what happened. The Committee on Publication Ethics (COPE) publishes a flowchart specifically for this, “Inappropriate image manipulation in a published article” (version 2, published 20 May 2024). It’s worth being precise about what this document is: it is a procedural decision tool for editors once a concern has already been raised, routing a case toward correction, an expression of concern, or retraction depending on whether the manipulation affects the paper’s conclusions and whether it appears intentional — it is not itself a technical method for detecting manipulation. Screening software and manual inspection (the steps above) supply the flag; COPE’s flowchart and editorial judgment decide what happens next. See CASRAI’s COPE guidelines explained for how that flowchart fits into COPE’s broader guidance.
What you raise should describe, specifically: which panel(s) or region(s), which category of concern (duplication, splicing, selective enhancement), and how you tested it (e.g. “overlaying panel 2B and 3A in Difference mode produces a near-exact match”). A specific, testable report is far more actionable for an editor than a general statement that a figure “looks manipulated.”
Frequently asked questions
Is image duplication in a research paper always misconduct?
No. ASM’s own pilot data above found that most flagged duplications were unintentional — typically an author reusing the wrong file when assembling a multi-panel figure — and were resolved through correction rather than any finding of misconduct. Intent, and whether the duplication affects the paper’s conclusions, are what separate an honest error from falsification under definitions like 42 CFR Part 93’s.
What counts as figure integrity beyond just avoiding duplication?
Figure integrity covers the full set of categories above — duplication, undisclosed splicing, and selective enhancement — plus disclosure: composite panels need a visible dividing line and a legend note, and any non-linear or selective processing needs to be described in the methods. A figure can be free of outright duplication and still fail an integrity check on disclosure grounds alone.
Is Western blot duplication treated differently from other image types?
Not procedurally — the same duplication/splicing/enhancement categories and the same overlay-based checking techniques apply. Western blots simply appear disproportionately often in image-integrity cases because they are assembled from multiple exposures/lanes more routinely than other figure types, which creates more opportunities for both accidental mix-ups and, less often, deliberate reuse.
Can I run a forensic check without buying commercial software?
Yes, for the manual techniques above — ImageJ/Fiji is free and handles overlay/subtraction comparisons, and free browser tools such as Forensically offer error-level-analysis and clone-detection passes. What free tools don’t replicate is the AI-assisted tools’ ability to check a submitted image against a large indexed database of previously published images (external duplication) — that requires the kind of reference index only a vendor like Proofig AI or ImageTwin maintains.
Related CASRAI resources
- Image manipulation — the operational definition and taxonomy this guide expands into a procedure
- Image duplication
- Scientific figure rules for journal submission — DPI, format, and submission specifications (a companion to this guide, not a substitute for it)
- Research misconduct: what an ORI finding actually means
- Falsification and fabrication
- Retraction and expression of concern
- Paper mill
- AI image-integrity screening goes mainstream: MDPI and ASM adopt Proofig, ImageTwin
- Dana-Farber $15 million False Claims Act settlement over image manipulation
- Northwestern $2.3 million NIH falsification settlement
Last verified 2026-08-16. Publisher-specific figure specifications and vendor adoption figures change; re-check the linked guide and news pieces directly if citing exact figures more than about 12 months from that date.
