Seeing Is Believing? Scientific Misconduct and the Detection of Problematic Images

One day in 2013, I did something that, in retrospect, changed the course of my career. I was reading something about plagiarism, and became curious to know if someone had ever plagiarized me. I copied a sentence from an article I had written in 2009 and pasted it into the Google Scholar search bar. To my surprise and anger, not only did my own paper appear, but an online book chapter did as well. Someone had copied three paragraphs I had written. That discovery led me down a rabbit hole of plagiarized sentences, paragraphs, and PhD theses. Although only two instances involved text I had written, I found dozens of other articles stealing from fellow researchers. Although my colleagues were watching Game of Thrones, I had found a new evening hobby alongside my daytime job as a microbiome researcher at Stanford University: investigating plagiarism. A couple of months later, a PhD thesis containing plagiarized text in the introduction led to a second serendipitous discovery. The thesis chapters had been published as peer-reviewed papers, and I noticed a Western blot panel with a distinctive black dot. I had seen that blot before in one of the other chapters. I checked, and it was indeed the same photo, but mirrored, stretched, and cropped differently. Other photos from the same thesis were also reused to represent different experiments. Combined with the plagiarism, this suggested an intention to mislead. I reported my concerns about the images in these papers to the respective journals, and both were eventually retracted. As with my first experience of being plagiarized, these discoveries were very disturbing. This was not science—this was cheating. Both findings confronted me with an alarming reality: the scientific literature, which I had always trusted, was not immune to misconduct. Science, for me, had always been about discovering the truth. Misconduct, such as plagiarism and the inappropriate reuse of images, is the opposite of what science should be about. It shows a blatant disregard for the truth.1 DISCOVERING SCIENTIFIC MISCONDUCT The discovery of errors or misconduct can happen at various stages of the scientific process. Sometimes, concerns are first raised by researchers who notice unusual behavior or questionable data from colleagues, even before a paper is submitted for publication. Issues may also be spotted during article review, when editors or peer reviewers detect plagiarism, anomalies, or ethical problems. However, peer review is not designed to catch fraud. Most reviewers assume the data are valid and may lack the time, tools, or expertise to spot manipulation. As the number of published papers, and thus submitted articles, continues to rise, the burden on journal editors and peer reviewers is growing.2 As a result, flawed papers can slip through, whereas rejected articles may be published elsewhere. Growing awareness has led some journals and publishers to adopt semi-automated screening tools for detecting plagiarism, image, or numerical irregularities.3,4 However, the most visible form of misconduct detection happens after publication, through post-publication peer review.5 Platforms like PubPeer,6 blogs, and social media now allow researchers to flag concerns in published papers, such as image duplication, suspicious statistics, or undisclosed conflicts of interest. A growing community of forensic metascientists—ranging from scholars to data analysts and retired researchers—routinely scrutinizes the literature for problems.7 This diverse group plays a crucial role in upholding scientific integrity. Many science “sleuths” have become specialized in finding specific types of data irregularities, such as plagiarism, unusual statistical distributions, overlapping or manipulated photographs, implausible plots, and problems with animal or human ethical approval. Recently, the forensic metascientist community released a set of open-access guides aimed at helping others detect problems in submitted and published papers. The Collection of Open Science Integrity Guides (COSIG) includes field-specific tips such as identifying problems in X-ray diffraction patterns or nucleotide sequences, as well as general guidance on how to best comment on PubPeer, how to spot ethical approval issues, or how to screen for plagiarism or image duplications.8 The Retraction Watch blog (https://retractionwatch.com/) and database (https://retractiondatabase.org/) complement PubPeer by tracking retractions, corrections, and expressions of concern, and providing background stories from affected authors and whistleblowers.9 However, science critics are not always welcomed; some have faced legal threats or harassment for raising concerns.10 Institutional investigations and retractions can take years to resolve, often delayed by conflicts of interest, reluctance to pursue high-profile researchers, or fear of litigation.9–12 IMAGE PROBLEMS Although many forms of sloppiness or misconduct cannot be identified by examining published papers alone, images are often the most visible form of data in which to detect problems. Photographic images, in particular, can be closely scrutinized for signs of inappropriate duplication with the unaided eye or by using software. After discovering those duplicated Western blot images in a PhD thesis, I began systematically examining biomedical papers for similar image problems by eye. I studied published photos of protein gels and blots, agarose gels, immunohistochemistry panels, fluorescent cells, mice, and plants. In total, I scanned 20,621 biomedical papers from 40 journals published between 1995 and 2014 for inappropriate image duplication.13 As I continued to identify more image copies, it became clear that problematic image duplications generally fall into three main categories (Figure 1):Figure 1.: Three categories of inappropriate image duplication. Each example was found as part of the study published in Bik et al.13 The images shown in this figure were originally published under a Creative Commons Attribution 4.0 International (CC BY) license; colored shape outlines were overlaid on the published images to highlight similar or duplicated regions. The source article DOI for each panel is indicated. The three examples illustrate (1) a simple duplication, where the same photo is visible twice, as illustrated by red boxes33; (2) a repositioned duplication, where multiple images, highlighted by boxes of the same color overlap,34 and (3) an altered photo, where multiple areas within the same panel have been duplicated, as indicated by circles and ellipses of the same color.35 All problems were addressed by the publisher by (C) a correction36 or (R) retractions.37 , 38 Simple duplications: These involve identical images reused to represent different experiments, such as repeated loading controls or microscopy images. Reuse is only considered problematic when the images are presented as distinct results. Simple duplications may result from unintentional errors, such as mislabeling a photo during capture or inserting the same panel twice in a figure collage. Duplications with repositioning refer to images that overlap or that are shown in a different orientation. Although a single overlap in the same orientation could be the result of mislabeling two photos taken from the same tissue sample, overlapping panels that have been stretched, rotated, or mirrored are more likely to be the result of an intention to mislead. Duplications with alteration involve the copying or rearranging of gel bands, lanes, cells, or other elements within an image. This category includes techniques such as stamping (the repeated use of the same area within an image) and patching, where parts of an image appear to have originated from elsewhere. Another notable subcategory is corner cloning, where small, repeated elements appear in the same corner of multiple panels within a figure. This may result from an author or journal staff member attempting to obscure a scale bar or label that did not conform to the journal’s house style. Although efforts to obscure a small smudge or an unwanted label might not be viewed as serious offenses, most manipulations in this category represent deliberate attempts to alter the visual record and may constitute scientific misconduct. To ensure that the duplications found in those 20,621 biomedical papers were not misinterpretations, all problematic figures were independently reviewed by two coauthors. A paper was classified as containing inappropriate figure duplication only if all three authors reached a consensus. In total, 782 (3.8%) of the 20,621 scanned papers were found to contain inappropriately duplicated figures. The incidence of such issues increased notably after the year 2000, from around 1% in 1995 to 2000, to over 4% from 2005.13 This rise in duplicated photos is likely linked to the shift from analog to digital photography around that time. Traditionally, images of blots and gels were taken by an institutional photographer and submitted as photographic prints alongside printed articles for peer review, which were then sent to the journal in an envelope. But as digital scanners became more common in labs, researchers began embedding photos directly into electronic articles submitted through email or websites. This made it easier to make errors as well as intentional changes. It was also observed that authors with one affected paper were more likely to have other problematic papers. Additionally, our analysis revealed substantial variation across journals, suggesting that editorial practices, such as pre-publication image screening, may play a crucial role in preventing these inaccuracies.13 It is important to note that not all image duplications are inappropriate. For instance, reusing the same loading control across multiple Western blot panels, or using the same lung tissue sample from a control animal in different treatment comparisons, can be scientifically valid, as long as the experimental conditions of the control sample were identical. Likewise, not all instances of inappropriate duplication constitute scientific misconduct. The three categories described above help distinguish varying levels of manipulation and gauge the likelihood of intent. Honest errors or sloppy figure preparation can result in simple duplications or overlapping photographic panels (Category I). If there is no intention to deceive and the authors can quickly provide the correct photos, such cases would not typically be considered misconduct. However, when a paper consists of multiple overlapping panels, duplicated photos that are rotated or mirrored (Category II), or photos that are manipulated (Category III), misconduct is much more likely. In our study, we found that at least half of the duplications bore signs of intentional duplication rather than honest error. Because our screening focused only on visible duplications in photographic images and did not include statistical inconsistencies, altered tables or graphs, or more sophisticated forms of falsification, the actual prevalence of misconduct is likely higher than the 2% we identified.13 HOW TO DETECT IMAGE PROBLEMS Initially, I screened biomedical papers by eye. I probably missed many duplications, especially those between papers rather than within a single article. It would be nice to be able to remember millions of images, but only very few people have such a gift. Today, a growing number of software tools and websites support scientific image forensics, each with their own strengths and limitations, and most image forensics specialists might use a combination of different tools.14 These tools can be roughly categorized into two main types: those designed for detailed graphical analysis of individual figures to detect image irregularities, and those capable of identifying duplications across multiple figures within a scientific paper or between papers. Software tools such as Mac’s Preview, Adobe Photoshop, GIMP, ImageJ, and Forensically allow for detailed analysis of individual images using features like contrast adjustment, overlays, and metadata inspection (Table, top half). Adobe Photoshop is widely used in formal investigations and is supported by toolkits and tutorials from organizations such as the U.S. Office of Research Integrity.15 Open-source alternatives, such as GIMP and ImageJ, offer comparable functionality. GIMP can be used for general image editing, and ImageJ for scientific applications, including plugins such as CLAHE (Contrast Limited Adaptive Histogram Equalization) to enhance local contrast and Look-Up Tables (LUTs) for false-color visualization.16,17 Forensically offers a set of free online tools to visualize images, and is helpful for detecting cloned regions within a single figure.18 However, it only works on duplicated elements in the exact same orientation and size, and may struggle with low-resolution or compressed images. Table. - Software for Image Forensic Analysis Tools for the analysis of a single figure Mac Preview GIMP Adobe Photoshop ImageJ/ Fiji Foto Forensics Forensically Free with macOS Free Licensed Free Free Free Simple contrast adjustment, add shapes, text boxes, etc. Advanced contrast adjustment, add shapes, text boxes, etc. Advanced contrast adjustment, add shapes, text boxes, etc. Droplets (automated sets of steps) available specific for image manipulation detection. Advanced image visualization, lookup tables for false coloring of black and white images; wide range of built-in and extensible plugins. Error level analysis and metadata inspection to reveal hidden edits, with color mappings that enhance subtle artifacts for easier detection. Error level analysis, Noise analysis, Principal component analysis, etc. Can detect cloned regions in the same orientation https://www.gimp.org/ https://www.adobe.com/creativecloud.html https://imagej.net/ij/ https://fotoforensics.com/ https://29a.ch/photo-forensics Software to screen PDFs or multiple images for image duplication or cloned regions ImageTwin Proofig ImaChek FigCheck Sherloq Licensed Licensed Licensed One free check per six months, or licensed Free Input: PDF, jpg, png, tif Input: PDF, jpg, png Input: PDF, jpg, png, tif Input: PDF, jpg, png, tif Input: jpg, png, tif, bmp, gif 100 million figures, user repository PMC figures, user repository User repository No library or repository No library or repository https://imagetwin.ai/ https://www.proofig.com/ https://www.imachek.com/ https://www.figcheck.com/ https://github.com/GuidoBartoli/sherloq The top part shows some popular tools that can be used for the analysis of a single figure, while the bottom part shows tools that can screen complete scientific papers as PDFs or multiple images to detect image duplication within or across figures. Reverse image search engines such as Google Lens, TinEye, and Yandex can be used to search the web for visually similar images or identical copies.19 These services can sometimes help identify recycled figures reused across publications or copied from vendor websites. However, they are generally not effective at detecting a single reused panel within a complex figure, nor can they access images from scientific papers behind paywalls. Specialized software tools such as the commercial packages ImageTwin, Proofig, and ImaCheck allow users to upload full PDFs or multiple images to detect direct, scaled, rotated, or mirrored image duplications, both within and between figures (Table, bottom half). Both ImageTwin and Proofig are in part powered by artificial intelligence (AI); they can also flag spliced blot panels and search for duplications across personal or large built-in or user-provided image repositories.20 Free alternatives like Figcheck and Sherloq offer more limited functionalities. Sherloq, an open-source forensic toolkit for Linux and Windows, provides basic duplication detection along with advanced tools such as luminance gradient visualization and metadata analysis, but is more complex to install and run.21 These tools are valuable resources for authors, publishers, and forensic metascientists at all stages of the publication process, as they can detect more image duplications than most individuals working by eye.14,22 They are increasingly being adopted as standard practice for screening incoming articles for image issues.4 However, these tools are not perfect. They may generate false positives or negatives, overlook subtle manipulations, and produce inconsistent results depending on image resolution, or For this review an part of the For the most should images, directly from the authors, that forensic tools rather than human As photos in scientific papers are the most and visible forms of data for detecting scientific misconduct. In high-profile in biomedical have involved image overlapping figure panels, or reuse of In Science has investigations that have patterns of image manipulation and misconduct across multiple In a Science revealed signs of image manipulation in a study that had the of The findings raised serious concerns that the study may have the for in both and the In the same the advanced an experimental while image were in papers by their main who was for data and fraud. these continued to both and the in In a and was the of a problematic images in more than of publications on and these cases highlight how problematic figures in scientific articles can to and on flawed or misconduct not only other researchers, who may and resources on false but also who may be by or from on honest In this misconduct both the of science and the of those who on The rise of a serious new for in the area of image integrity. Although tools can offer such as helping with scientific articles in they can also be to produce photos, graphs, or experimental that appear often the and of scientific to detect with the eye. This concerns for as visual data could inspection and that a was an photo or it but that no with can generate detailed and images of people and complete with subtle and If they can such with complex and scientific images, such as Western blots or images of and even often contain distinctive features that would allow a eye to identify images to detect specialized forensics The software described in the part of the are designed to detect figure duplications, but photos are they cannot be by such concern, Photoshop now with features such as where a user-provided photo can be in one or multiple by thus new and panels or areas that could be used for in scientific images can (Figure of being used to a panel from panels from of et by the journal in used as for to an panel by Photoshop is widely available and to even individuals with or no experience can quickly generate This the for scientific images, detection and even more The by images the for more image screening editorial and awareness of the and of in the scientific process. In an the of between and data to in the scientific of on detecting photos, could images as while editorial might of image for articles containing Open like for and by the and supported by and embedding into media These of and was as a for digital might be used in or scanners used in to ensure the photos they produce are However, scientific and publishers In not To the of the scientific be with new practices, and like should be and as part of and of scientific misconduct is and some is being are increasingly tools such as image analysis plagiarism screening, statistical and paper to screen Many now data and study to make it to For journals could all experimental data a submitted article to be in an online not only for the data in the but for experimental this for honest researchers and upload their data it also a substantial of for individuals or organizations But such are not on their As seen by the rise of scientific are support and most from publication and over and these even the best tools fall misconduct is not a new but in and is more images, and publications can not only but also and the some are attempting to the of misconduct and the of to science that not with their misconduct is in this to on or or best of science is and a to To we peer review, by tools, data and a that over or in science on both and then can science to help the truth. of The author has and from publishers and and through free access to ImageTwin and Proofig software. is an of the at Stanford This was is for and for the Photoshop example in The author using and for text in the preparation of this article. that the and of the during and that had no in the of this The author full for upholding the of the and presented in this article.