Scientific Image Integrity in Medical Journals: What Authors Need to Know in 2026

For most of the history of biomedical publishing, figure integrity depended on what a journal editor and an attentive reviewer could catch by eye. That changed when Elisabeth Bik and colleagues published a systematic screen of over 20,000 biomedical papers in 2016 and found problems in 3.8% of them, with at least half of those cases showing signs of deliberate manipulation. The study did not shake the foundations of science overnight, but it shifted editorial thinking. Publishers began to ask whether manual review was ever going to be enough.

By 2025 and 2026, several major publishers had their answer. The American Association for the Advancement of Science (AAAS), which publishes Science and five other titles, adopted Proofig AI for image integrity screening. The American Society for Microbiology (ASM) completed a 12-month pilot with Imagetwin and made it a permanent part of its ethics workflow. MDPI, the world's largest open-access publisher by article volume, announced a multi-year deal with Proofig following an internal evaluation that tested every available tool. These are not fringe publishers. If you are preparing a biomedical manuscript today, there is a real chance your figures will be algorithmically screened before anyone reads your conclusions.

Why Image Problems Are So Common in Biomedical Research

The honest answer is that figure assembly is tedious, happens late in the manuscript process, and often falls to whoever is least senior on the team. A Western blot experiment generates multiple exposures. A confocal microscopy session produces hundreds of image files, many of them visually similar. When researchers compile final figures weeks or months after collecting data, honest mistakes happen. A technician copies the wrong lane from an archived folder. A panel from a pilot experiment appears in the final figure alongside data from the formal study. Two figure panels end up showing the same cells under different conditions, then get used again in a supplementary figure without the authors noticing.

These errors sit alongside a smaller category of deliberate fabrication, and the difficulty for both editors and automated tools is that the same artifact can result from carelessness or fraud. What has changed is the ability to find both. Proofig and Imagetwin detect duplications, rotations, flips, scale distortions, and region splicing regardless of intent. When a screen flags something, the journal asks authors to explain it or provide original files. If neither is possible, the result is the same whether the original mistake was accidental or intentional.

One specific driver of the problem that often goes unmentioned: data archiving practices in academic labs. When researchers leave a group, they typically hand over a folder of experimental data, but that folder is rarely organized in a way that maps cleanly to the published figures. When the first author on a paper moves to a new institution, the ability to retrieve the original raw image for panel 3B of Figure 2 becomes genuinely difficult. Publishers are increasingly aware of this, which is why several, including journals in the Nature Portfolio, now ask for raw or minimally processed versions of key images at the time of submission rather than only when post-publication questions arise.

What AI Screening Tools Actually Detect

Proofig and Imagetwin operate on related but distinct principles, and understanding what they look for helps authors prepare. Both tools compare image regions within a manuscript and, in the case of Proofig, also run comparisons against images in PubMed and other public databases. This means a figure that was legitimately published in a prior paper and then reused without attribution can be detected even if the original manuscript predates the current submission by years.

Within a single manuscript, the tools flag regions that appear in more than one location. Critically, they are robust to the kinds of transformations that authors have historically used to disguise reuse: rotating an image 90 degrees, flipping it horizontally, resizing it, adjusting brightness or contrast globally, or changing the color channel. A Western blot lane that appears in Figure 1C, then again with slightly different brightness in Figure 4A, will still be matched. The same applies to fluorescence microscopy images where a cell cluster is copied from one figure panel and presented as a different experimental condition in another.

Proofig added Western blot splicing detection in late 2023, which addresses a different and particularly common problem in molecular biology. Splicing means taking bands from two or more separate blot runs, or from two positions on the same blot, and presenting them as a single continuous image. This practice is sometimes done legitimately to show lanes that were not originally adjacent, but current journal policy across the Nature Portfolio, ASM, Elsevier, and others requires any spliced blot to be clearly marked with a dividing line in the figure and explained in the legend. When that disclosure is missing and a tool identifies the splice, it is treated as a potential integrity violation.

The PubMed cross-reference check is worth underlining because many authors do not expect it. If you have previously published a paper with a histological image of tissue from a mouse model, and your new manuscript includes a figure from what you describe as a different cohort but the images are visually identical, the tool will return a match. This is one of the most common legitimate problems that journals encounter: a researcher uses a representative image from a prior study as a “typical example” in a new paper, believing it is close enough in experimental conditions to be illustrative, without realizing that reuse of published images in a new context requires explicit disclosure and, at many journals, permission from the original publication.

Which Journals Are Now Running These Screens

The adoption curve is steeper than most authors realize. AAAS began rolling out Proofig across its six journals (Science, Science Advances, Science Immunology, Science Robotics, Science Signaling, and Science Translational Medicine) in early 2025. These journals collectively publish several thousand papers per year, mostly in life sciences and clinical research. By mid-2025, image screening was integrated into the post-revision workflow: manuscripts that had cleared peer review were run through Proofig before acceptance was confirmed.

ASM's integration of Imagetwin covers all of its journals, a substantial portfolio that includes mBio, the Journal of Bacteriology, Infection and Immunity, Journal of Clinical Microbiology, and several others. The 12-month pilot that ran in 2023 and 2024 screened over 2,600 manuscripts and has now transitioned to a permanent ethics checkpoint. ASM has been explicit about one finding that should reassure authors: most flagged issues were unintentional and the majority of those were resolved when authors submitted original images. The 0.23% revocation rate means that having a figure flagged does not automatically mean rejection. What it means is that you will need to respond quickly with documentation.

MDPI's deal with Proofig is significant partly because of scale. MDPI publishes across a very wide range of biomedical disciplines, from oncology to biomolecules to nutrients to viruses, and has historically been a target of criticism over quality control. The multi-year Proofig agreement follows a prolonged internal evaluation of available tools and represents a structural commitment to pre-publication screening rather than relying on post-publication scrutiny. The rollout started with selected biomedical titles and is intended to expand across the full portfolio.

Beyond these specific announcements, the broader trend is clear. The Journal of Clinical Investigation has discussed Proofig in the context of strengthening scientific integrity. Multiple Nature Portfolio journals have maintained detailed image integrity policies for years and use similar checking processes internally. Cell Press titles, Elsevier specialty journals, and BMJ journals all have written image manipulation policies and use some form of image review. The specific tools vary, but the direction is consistent: automated screening before publication is becoming the norm rather than the exception.

The Rules on Acceptable and Unacceptable Manipulation

Authors often find the distinction between legitimate processing and prohibited manipulation confusing because the rules are genuinely contextual. The principle across Nature Portfolio journals, which is representative of broad industry consensus, is that processing is acceptable when it is applied uniformly to the entire image and does not obscure, eliminate, or misrepresent any part of the original data.

Adjusting brightness and contrast globally across an entire image is generally acceptable, provided the adjustment is the same for every part of the image and does not cause data to disappear from the visible range. The same adjustment applied to one region and not another is not acceptable. A brightness increase that reveals detail in a dark area is legitimate if it is applied uniformly. A brightness increase applied only to a specific band in a Western blot to make it appear stronger than it was in the original exposure is not.

Pseudocoloring, which means assigning color channels to fluorescence microscopy data that was acquired in grayscale, is standard practice and acceptable, but it must be disclosed in the figure legend with an explanation of which color corresponds to which fluorescent marker. Gamma correction, which alters the relationship between signal intensity and pixel brightness, is acceptable at some journals if explicitly noted, but is prohibited or restricted at others. Nonlinear adjustments in general require disclosure at Nature Portfolio and are treated with suspicion at many clinical journals where the expectation is that figures represent the data as closely as possible to how it was captured.

Western Blots and Microscopy Images: Where Most Problems Arise

Western blots are the most common source of image integrity concerns in biomedical journals, partly because of how they are produced and partly because of how researchers have historically treated them in figures. A blot run involves multiple lanes on a single membrane, and it is technically common to cut gels and membranes, to strip and reprobe, or to run multiple proteins across separate blots. When the final figure is assembled, lanes from different runs or different positions on the same blot are frequently placed side by side.

Most journals now require that any such rearrangement be marked with a clear dividing line between the juxtaposed sections, noted in the figure legend, and supported by the availability of the full original blot image. This requirement has been strengthened by the realization that some researchers were removing the dividing lines after assembling the figure, either accidentally when cropping or deliberately to create a cleaner visual. AI tools that detect splicing look specifically for the telltale pixel discontinuities or boundary artifacts that appear at splice sites, even when visual dividing lines have been removed.

For fluorescence microscopy, the most common problems are copying a field of view from one condition and using it to represent another, and applying differential processing to individual color channels in a merged image without disclosure. The second of these is often genuinely unintentional. When a merged image looks unbalanced, a researcher may adjust the green channel independently to improve visual clarity, which is sometimes acceptable but must be stated explicitly. When the channel adjustment changes the apparent colocalization of two markers, which is often the primary scientific conclusion being illustrated, the adjustment is no longer cosmetic and becomes substantive manipulation.

Flow cytometry plots, histology sections, and electron microscopy images each carry their own version of these problems. For histology in particular, one issue that has been documented repeatedly is the use of a single tissue section photographed at different magnifications to represent what is described as tissue from different animals or different treatment groups. The Proofig cross-reference against PubMed images is especially effective here, because archival histology images are visually distinctive and hard to disguise.

How to Prepare Your Figures Before Submission

The most effective thing you can do is treat figure preparation as a documentation exercise rather than a design task. Every panel in every figure should have a corresponding source file, and that source file should be traceable to an experiment in your lab notebook or data management system. This is not only good practice; it is increasingly a formal requirement. Several journals in the Nature Portfolio and elsewhere now request original, minimally processed image files at submission or revision. If you cannot locate the source file for a figure panel, that is a warning sign worth addressing before submission rather than after.

When you assemble a multi-panel figure, keep a record of which source file corresponds to which panel. Label them clearly and keep the record accessible to all co-authors. The corresponding author is responsible for responding to editorial queries about image integrity, but may not have been the person who ran the assay or assembled the original figure. In practice, many integrity issues come to light when a journal asks for source files and no one on the team can locate them, not because the data was fabricated but because archiving was haphazard. Good file documentation resolves that problem before it becomes a crisis.

For Western blots, the practical standard is to save the original scan of the entire membrane, not just the cropped region that appears in the figure. Most journals that require blot submission want the full membrane image with clear indication of where the displayed region sits within it. The molecular weight markers should be visible, the loading control lane should be present, and if multiple antibodies were used on the same membrane after stripping, each probing should be documented separately. Some groups photograph their blots with a ruler or timestamp to create an unambiguous archive. This sounds burdensome, but it takes minutes and eliminates a category of editorial query that can otherwise take weeks to resolve.

When an AI Screen Flags Your Figures

The ASM pilot data offers a useful calibration. Of 2,627 manuscripts screened, 3.9% had flagged concerns. Most were resolved when authors provided original images and explanations. Acceptance was revoked in only six cases, all involving issues that persisted after authors had a chance to respond. This means that the majority of authors who receive an image integrity query from a journal will be in a position to respond satisfactorily if they have their source files.

What helps most in responding is speed and specificity. When a journal contacts you about a flagged figure, they will typically describe the specific panels involved and the nature of the concern. Responding with the original source files and a clear explanation of the experimental workflow is much more effective than a general statement that the data are authentic. If the flagged issue reflects a genuine error in figure assembly, acknowledge it, provide the corrected figure with documentation, and explain how the error occurred. Journals are generally accustomed to these kinds of assembly errors in good-faith manuscripts.

What escalates a query into a retraction or acceptance revocation is usually one of three things: the authors cannot produce source files to support the figure, the explanation provided is inconsistent with the flagged image (for example, claiming an adjustment was applied uniformly when the pixels show it was not), or additional problems surface when the editorial team looks more closely at the full manuscript. The first of these is preventable with good archiving. The second is preventable with honest communication. The third is harder to control once a manuscript is under scrutiny, which is another reason to prepare figures carefully from the start.

The Broader Shift Toward Pre-Publication Screening

Image screening sits alongside a set of other automated checks that are becoming part of standard editorial workflows at major journals. Plagiarism detection via iThenticate or Turnitin has been in use for over a decade. Citation verification tools that check whether references exist and are accurately described are being tested at several publishers. Statistical review tools that flag anomalous data distributions, impossible values, or inconsistencies between text and tables are in early adoption at some clinical journals. The cumulative effect is that the post-acceptance period, from conditional acceptance to formal publication, is becoming significantly more demanding for authors than it was five years ago.

Authors who understand this shift before they submit are in a much better position than those who encounter it mid-process. If you know that your target journal uses image screening, you can run your own figures through a tool like Proofig before submission. Proofig offers a researcher-facing product that allows authors to pre-check their own manuscripts against the same kind of detection that journals use. Running this check before submission does not guarantee a clean editorial screen, since cross-database comparisons against PubMed require access to the broader database, but it will catch within-manuscript duplications and the most common assembly errors.

The deeper point is that image integrity has moved from being a problem that surfaces occasionally in post-publication review to being a formal submission requirement. The tools exist, the publishers are deploying them, and the workflow now includes an integrity checkpoint between peer review and publication that did not formally exist for most journals five years ago. Treating figure preparation with the same rigor you apply to statistical analysis is no longer exceptional practice. It is the current floor.

What This Means for Lab Culture

The image integrity shift has practical implications for how lab groups organize their data and who is responsible for figure quality. In many research groups, figure assembly is treated as a late-stage task delegated to whoever has the software skills, rather than as a scientific step with its own accountability. That model is no longer compatible with the editorial environment being described here.

The most durable approach is to establish figure documentation standards at the time of data collection, not at submission. When an experiment is run, the original output files should be archived immediately with clear naming that links them to the protocol and date. When a figure is assembled, a companion record should note which source files were used for each panel and what processing was applied. This record does not need to be elaborate. A shared spreadsheet with one row per panel, listing the source file path, acquisition date, instrument, and processing steps, takes fifteen minutes to set up and can prevent weeks of recovery work when an editorial query arrives.

For principal investigators, the practical implication is that data management training and figure documentation standards should be part of lab onboarding, not something left to students to figure out independently. Several institutions have updated their research integrity training materials in the past year to include specific guidance on figure preparation in light of the new screening tools. If your institution has not, the COPE guidelines, the Committee on Publication Ethics resource pages, and the Nature Portfolio's image integrity documentation are all publicly available and serve as reasonable starting points.

Further Reading