03 · What You Need to Know
How to Distinguish Defensible Image Processing From Misleading Alteration
What Is Legitimate Scientific Image Processing?
Scientific image processing involves computational transformations applied to images or imaging measurements to support visualization, reconstruction, analysis, or communication.
Examples include calibrated intensity adjustments, image registration, background correction, reconstruction from instrument measurements, and certain filtering procedures.
These operations can be legitimate when they are appropriate for the scientific purpose, preserve the relevant information, and are documented sufficiently for evaluation.
For example, adjusting the brightness and contrast of an entire microscopy image may improve visibility without necessarily changing its substantive interpretation. However, excessive contrast adjustment could eliminate weak signals or exaggerate differences between experimental conditions.
The U.S. Office of Research Integrity (ORI) emphasizes that scientific images should be treated as data and that original image files should be preserved. Nature Portfolio likewise requires published images to represent the original data accurately and conform to disciplinary standards.
Thus, the acceptability of processing depends on its actual effect, not merely the name of the software operation.
What Is AI Image Manipulation?
AI image manipulation broadly refers to modifying images through artificial intelligence methods. The term itself is not necessarily synonymous with misconduct.
For example, an AI model may perform image segmentation, denoising, reconstruction, or correction of acquisition artifacts as part of an established analytical workflow.
Problems arise when the transformation changes scientifically meaningful information without adequate justification or presents model-generated content as though it were directly observed.
A general-purpose generative editor might remove a visible structure, fill a missing region, or synthesize plausible texture. These capabilities may be useful for creative illustration but can be inappropriate when applied to empirical research images.
Researchers should therefore distinguish the technical act of modifying an image from the scientific question of whether the modification preserves evidential fidelity.
Legitimate image processing
A documented transformation that is scientifically justified, appropriately validated, and consistent with faithful representation of the research evidence.
Misleading image manipulation
An alteration that misrepresents what was observed or measured, including through selective removal, invented features, or distortion of scientifically relevant information.
Which Scientific Image Adjustments Are Generally Acceptable?
Image integrity guidance from ORI and Nature Portfolio recognizes that some adjustments may be appropriate when they do not distort the underlying evidence.
However, acceptability depends on the imaging modality, intended analysis, and journal requirements.
| Processing Operation |
Potentially Legitimate Use |
When It Becomes Problematic |
| Brightness adjustment |
Improve visibility through consistent adjustment of the image. |
Concealing weak signals or altering comparisons between experimental conditions. |
| Contrast adjustment |
Display relevant intensity differences without losing meaningful information. |
Clipping data, eliminating background structures, or exaggerating apparent differences. |
| Cropping |
Focus attention on a relevant region while preserving interpretive context. |
Removing contradictory evidence or changing the apparent relationship between structures. |
| Image registration |
Align images for justified comparison or analysis. |
Applying undocumented transformations that alter spatial measurements or create misleading composites. |
| Background correction |
Correct a characterized acquisition-related background signal. |
Removing genuine biological or physical information as though it were background. |
| AI denoising |
Reduce acquisition noise using a method validated for the intended endpoint. |
Suppressing weak real signals or introducing unsupported features. |
| AI super-resolution |
Estimate higher-resolution representations for an appropriately validated analytical purpose. |
Presenting inferred fine details as independently observed structures. |
| Object removal or generative filling |
May have legitimate uses in explicitly synthetic materials or specialized validated methodological research. |
Removing observed structures or inventing missing content in an image presented as authentic evidence. |
This table is a practical interpretation of image integrity principles, not a universal permission list. A journal may prohibit a particular technique even when a researcher believes its effects are scientifically harmless.
Why Is Applying Adjustments to the Entire Image Important?
Selective processing can create misleading differences between regions of the same image.
For example, increasing the contrast around a particular cell while leaving surrounding cells unchanged may exaggerate its apparent importance. Selectively darkening a background region could conceal a weak signal that complicates the interpretation.
Nature Portfolio advises that brightness and contrast adjustments should generally be applied across the entire image and equally to controls. Its guidance also discourages processing that emphasizes one region at the expense of another.
ORI similarly identifies selective enhancement as a questionable practice, particularly when it changes the apparent relationship between observed features.
However, whole-image processing is not automatically valid. An excessively aggressive global adjustment can still eliminate meaningful information.
The scientific effect of the transformation remains the decisive consideration.
Can Cropping a Scientific Image Be Considered Manipulation?
Cropping is commonly used to focus attention on relevant information, but it can also change the context of an observation.
For example, cropping a microscopy image to highlight a particular structure may be appropriate when the cropped region remains representative of the observation.
By contrast, cropping away neighboring structures that contradict the proposed interpretation may mislead readers.
In electrophoretic gels and blots, removing lanes or presenting nonadjacent lanes as though they were continuous can create additional integrity concerns.
Nature Portfolio requires rearranged gel lanes to be clearly indicated and described. The original image should remain available for verification.
Cropping is therefore not inherently inappropriate, but its consequences for interpretation must be examined.
Why Is Removing an Object Different From Correcting an Acquisition Artifact?
An acquisition artifact arises from the imaging process rather than the phenomenon being investigated. Examples may include certain detector defects, motion-related distortions, or known optical effects.
Correcting a documented artifact using a validated procedure may be scientifically appropriate.
However, an object that appears unwanted is not necessarily an artifact. It may represent genuine information relevant to the study.
For example, a researcher may regard a faint band in an experimental image as contamination. Removing it with an AI editing tool would change the visual record unless the operation is part of a justified analytical procedure that does not misrepresent the observation.
ORI discourages cloning and selective object removal in scientific images because these operations can change both the content and context of the evidence.
The detailed question of using AI to remove objects or artifacts from scientific images requires establishing what the feature represents before deciding whether any correction is justified.
Can AI Fill Missing Image Regions Without Falsifying the Evidence?
Generative inpainting can synthesize plausible content for missing or obscured regions. The resulting pixels may resemble surrounding structures without being directly supported by observations of the missing area.
This can be useful in explicitly synthetic illustrations or certain methodological investigations where generated regions are clearly identified as estimates.
It becomes problematic when reconstructed content is presented as though it were observed in the original experiment.
For example, filling a missing portion of a microscopy image with generated cellular structures could introduce features that were never recorded.
Researchers should distinguish estimated or simulated information from authentic observations and avoid using unsupported generated details as empirical evidence.
The risks associated with filling missing scientific image details are particularly serious when the generated region influences measurements or conclusions.
Can AI Denoising and Upscaling Count as Legitimate Scientific Processing?
Potentially, but both require validation appropriate to their scientific use.
Denoising models may suppress weak genuine signals because those signals resemble noise. Super-resolution models may estimate fine structures using patterns learned from training data.
Neither operation is automatically equivalent to recovering information that was directly measured.
Researchers should examine the effects of AI denoising on meaningful signals and the possibility that AI upscaling introduces unsupported details.
Where enhanced images support quantitative measurements, validation should address the actual measurement endpoint rather than relying only on improved visual appearance.
Scientific legitimacy depends on what the method can reliably establish under the relevant acquisition conditions.
Does Using the Same Processing Settings Make an Image Comparison Valid?
Consistent processing is important when images are compared, but it is not sufficient by itself.
For example, applying the same contrast settings to treatment and control images reduces the risk of selectively emphasizing one group.
However, a nonlinear AI model may respond differently to images with different noise levels or structural characteristics, even when identical settings are used.
Researchers should evaluate whether the processing introduces systematic bias or changes the quantities being compared.
For quantitative measurements, calibration and appropriate validation are necessary.
The principle is that processing should not create apparent experimental differences that arise primarily from the transformation rather than the original observations.
How Do Journal Policies Distinguish Processing From Manipulation?
Journal policies commonly emphasize faithful representation of original data, transparency about processing, and preservation of unprocessed images.
Nature Portfolio states that submitted images should be minimally processed and that final images must accurately represent the original data. Its guidance addresses contrast adjustments, microscopy processing, gel and blot presentation, and documentation of analytical procedures.
Nature's research figure guidance also explicitly restricts generative AI use in figures, including content-aware editing tools. This restriction is more specific than the general methodological question of whether an AI-based image processing technique can be scientifically validated.
Researchers should therefore distinguish scientific defensibility from editorial permission.
A method may be useful in a research workflow but still be prohibited for a particular form of published figure.
Before submission, researchers should check the current policies of their target journal rather than assuming that all publishers apply identical standards.
Why Must Researchers Preserve the Original Image?
The original image provides the reference against which subsequent processing can be evaluated.
Without it, researchers may be unable to determine whether a feature was observed, enhanced, suppressed, or generated during processing.
ORI recommends retaining unprocessed image data and performing modifications on copies. Nature Portfolio also advises retaining original image files and metadata because editors may request them during manuscript review or investigations of published work.
Researchers should preserve acquisition information and processing records sufficient to reconstruct the transformation where feasible.
The importance of preserving original scientific images is therefore directly connected to the credibility and verifiability of the research record.
Can Inappropriate Image Manipulation Constitute Research Falsification?
Yes. Image manipulation may constitute falsification when research data or results are changed or omitted such that the research record no longer accurately represents the work.
However, not every inappropriate image-processing decision automatically constitutes formal research misconduct.
Under the U.S. federal research misconduct framework, a finding of misconduct requires additional elements, including the applicable standard concerning intent, knowledge, or recklessness. Honest error is distinguished from misconduct.
ORI cautions that identifying an image discrepancy does not by itself establish falsification or intent to deceive. Further investigation may be necessary.
Researchers should therefore distinguish a scientifically invalid processing method, a questionable reporting practice, and a formally established misconduct finding.
The separate question of when AI image enhancement crosses into research falsification depends on the nature and consequences of the alteration and the applicable integrity standards.
Watch Out
Disclosure does not make a misleading image acceptable. A researcher cannot justify removing observed features, inventing missing structures, or exaggerating experimental differences merely by stating that AI processing was used.
How Can Researchers Determine Whether Processing Is Defensible?
A defensible evaluation should consider the original image, the purpose of processing, the transformation applied, and its effect on the scientific conclusion.
Researchers can ask whether the operation is necessary, whether it has been validated for the imaging modality, and whether it changes any feature that matters to the research question.
They should also consider whether the procedure is reproducible and whether the published representation accurately communicates what was measured or inferred.
Where the image is used for quantitative analysis, the effects of processing on measurements should be assessed directly.
For complex or model-based procedures, independent reference data and appropriate sensitivity analyses may be needed.
Scientific image integrity is not established by the absence of visible editing artifacts. It depends on the evidential fidelity of the final representation.
04 · A Practical Example
Comparing Legitimate Contrast Adjustment With Misleading AI Object Removal
Hypothetical Example
Two Researchers Prepare Microscopy Images for Publication
Two researchers acquire fluorescence microscopy images to examine the distribution of a cellular marker.
Both researchers find that the original images are difficult to interpret because of low contrast and background variation. They apply different processing approaches before preparing their manuscript figures.
This example is hypothetical and illustrates image integrity principles rather than an actual misconduct case.
Researcher A: Documented Whole-Image Adjustment
The first researcher applies a modest, documented contrast adjustment across the entire image using the same procedure for comparable experimental conditions.
The adjustment improves visibility without clipping meaningful intensity ranges or concealing observed structures.
The original images are preserved, and the processing procedure is documented.
Subject to the relevant imaging and journal requirements, this may constitute legitimate image processing.
Researcher B: AI Removal of an Inconvenient Structure
The second researcher notices a faint fluorescent structure that complicates the proposed interpretation.
Instead of investigating whether it represents genuine signal or an acquisition artifact, the researcher uses a generative editing tool to remove the structure.
The edited image appears cleaner and more consistent with the expected result.
However, the final image no longer faithfully represents the recorded observation.
Compare the Two Processing Decisions
| Criterion |
Researcher A |
Researcher B |
| Purpose |
Improve visibility of recorded information. |
Remove a feature that complicates interpretation. |
| Transformation |
Documented whole-image contrast adjustment. |
Selective generative object removal. |
| Effect on evidence |
Relevant features remain represented. |
An observed feature disappears. |
| Scientific justification |
Processing is assessed for fidelity and consistency. |
No evidence establishes that removal is a valid correction. |
| Integrity assessment |
Potentially legitimate, subject to validation and policy. |
Misleading alteration of the research record. |
What Should Researcher B Have Done?
The researcher should retain the original image and investigate the nature of the unexpected structure.
If the feature is a documented acquisition artifact, an appropriate correction procedure may be considered, subject to validation and reporting requirements.
If it represents genuine signal, it should not be removed merely because it conflicts with the expected interpretation.
If its origin remains uncertain, the uncertainty should be addressed through additional observations, analysis, or appropriate qualification.
The example illustrates why the scientific status of an image alteration depends on what it does to the evidence, not simply whether the software can perform it convincingly.
