> For the complete documentation index, see [llms.txt](https://user-manual.imagetwin.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://user-manual.imagetwin.ai/detailed-image-analysis/detailed-image-analysis-1/figure-screening-tips-from-an-image-integrity-expert.md).

# Figure Screening Tips from an Image Integrity Expert

Jana Christopher on what a clean figure should never show, and how to catch the ones that aren't.

Image screening is often mistaken for an intent test, is this fabrication or not? In practice, most problems aren't fabrication but rather figures that are technically honest but impossible to interpret correctly, alongside the smaller share that are altered on purpose. Both need the same close look.

#### <mark style="color:$primary;">**What Makes Figures Hard to Screen**</mark>

Unlike text, an image carries information in places that aren't obvious at a glance: acquisition settings, processing history, background noise, channel data. Two figures can look identical and mean completely different things, or look completely different and share the same underlying data. That's what makes screening a skill, not a glance.

#### <mark style="color:$primary;">**What a Clean Figure Should Never Show**</mark>

Before anything else, check the basics:

* Image intensities being compared without identical acquisition and processing settings

<figure><img src="https://4155828172-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FoBwXHdADuISE0Cd1Ah08%2Fuploads%2FJoNeNNAgYh9bSCXzYn7g%2FScreenshot%202026-09-21%20at%2014.54.30.png?alt=media&amp;token=383ce77a-0d53-4a7c-ad3e-244bb902599e" alt=""><figcaption><p>The same microscopy image adjusted at two different brightness/contrast levels, resulting in two different vesicle counts.</p></figcaption></figure>

* Missing, wrong, or badly placed scale bars, especially for tumors, plants, or archaeological specimens where a ruler in the shot beats a scale bar next to it
* Color scales that start and end on the same shade, hiding the difference between the top and bottom of a structure
* Color choices that aren't colorblind-safe, since red-green color blindness affects roughly 8–10% of men and can make a figure unreadable to a meaningful share of readers

<figure><img src="https://4155828172-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FoBwXHdADuISE0Cd1Ah08%2Fuploads%2FIGeZOpqeXYPX1rGhcxFZ%2FScreenshot%202026-09-21%20at%2014.55.19.png?alt=media&amp;token=2adc40b4-b824-48c5-9c91-c408be6625f0" alt=""><figcaption><p>The same cell image as it looks with normal color vision vs. simulated red-green color blindness. The colors become indistinguishable.</p></figcaption></figure>

#### <mark style="color:$primary;">**A Repeatable Screening Routine**</mark>

1. Scan every figure by eye for anything that looks off.
2. Group images of the same type (all western blots, all microscopy panels) and tile them side by side to compare directly for duplication.
3. Look for alterations, doctoring, and processing irregularities such as over-contrasting.
4. Separate the color channels in multi-channel or merged images, sometimes only one channel has been edited, or the channels don't line up.
5. Check scale bars, magnification, and whether magnified insets are correctly marked in the source panel.
6. Check whether bar graphs, plots, and charts are plausible, and look for duplicated or repeating patterns.
7. Check animal welfare compliance, such as tumor volumes against ethical limits.

#### <mark style="color:$primary;">**Catching Deliberate Manipulation**</mark>

Solarizing an image is one of the fastest ways to expose a doctored region: cloned or copy-pasted background patches that look seamless at normal contrast become obvious once solarized, because the underlying noise pattern matches exactly in a way that doesn't happen by accident. This is the same technique that first uncovered a set of fabricated western blots submitted in 2018, later traced across dozens more submissions.

<figure><img src="https://4155828172-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FoBwXHdADuISE0Cd1Ah08%2Fuploads%2FPwjPxFYAmUE9TgtLXoK9%2FScreenshot%202026-09-21%20at%2015.01.55.png?alt=media&amp;token=0b775dfb-74b8-43e1-bcf4-dca55695339c" alt=""><figcaption><p>A gel image after solarizing, with two cloned/copy-pasted background patches highlighted in yellow.</p></figcaption></figure>

Where raw data is available, it usually settles the question outright: original files and their metadata (acquisition timestamps, software history, camera and instrument details) either support an author's explanation or contradict it.

<figure><img src="https://4155828172-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FoBwXHdADuISE0Cd1Ah08%2Fuploads%2FnZ2zEQrPBs7QEI0NCSq7%2FScreenshot%202026-09-21%20at%2015.04.41.png?alt=media&amp;token=5471276a-c8a4-477a-923c-0c8d8b48a104" alt=""><figcaption><p>A raw image metadata panel showing acquisition date, instrument settings, and edit history, the data behind the figure.</p></figcaption></figure>

#### <mark style="color:$primary;">**Field Tips**</mark>

* Composite images assembled from multiple source files often give themselves away through hard edges around a specific region, like a set of nuclei that doesn't match its surroundings.
* Duplicated citations don't stop at retraction. Papers with unresolved image issues can keep accumulating citations for years after being flagged or retracted, which is part of why catching problems before publication matters as much as it does.
* Most of the job isn't making a final call on intent. It's asking the right questions and flagging them for someone who can.