A custom AI reviewer for image quantification integrity

PerfectPaper’s image quantification agent reads a manuscript for the gaps behind a microscopy figure: how many fields and how many animals were sampled, how fields were chosen, whether scoring was blinded, how the threshold was set, and whether the representative image shown is doing the work a quantification should be doing.

Copy the brief below into a custom agent slot, and enable vision so it reads the panels rather than only the captions.

The recurring problem is not fabrication. It is that a beautiful representative image sits where a quantification should be, and no one states how many fields, sections or animals stand behind it. Fabrication has been measured and it is the smaller problem: Bik, Casadevall and Fang screened 20,621 published papers and found 3.8% contained problematic figures, with at least half showing features suggestive of deliberate manipulation (mBio 2016;7(3):e00809-16). Undescribed sampling is not 3.8% of manuscripts. It is closer to the default, and it is invisible because nothing about the figure looks wrong.

When to use this agent

• Your manuscript quantifies immunohistochemistry, immunofluorescence, or histology
• A central claim rests on a representative image panel
• Scoring was done by eye, or by a threshold chosen while looking at the result
• Fields were selected by the person who knew the treatment group
• You use ImageJ or a similar tool and the macro is not described
• Your figure legend reports a mean and an error bar but no unit of analysis
• You compare fluorescence intensity between groups imaged on different days
• A reviewer has asked what “positive” means in your positivity index

The agent brief

Paste this into a custom agent. Suggested settings: work type figures_tables, skill level graduate, capability vision.

Name: Image quantification integrity

You are reviewing microscopy and histology quantification. Read the methods,
figure legends, and the figures themselves where you can see them.

Determine and report:

1. Quantification versus illustration. For every claim supported by an image
   panel, determine whether an accompanying quantification exists. Report every
   claim of a difference between groups supported only by representative
   images. This is the most important item; report it first.

2. Sampling. Determine how many fields per section, how many sections per
   animal or sample, and how many animals or samples contribute to each
   quantification. Report each number that is absent. A figure legend giving
   only one of these three is incomplete; say which are missing.

3. Field selection. Determine how fields were chosen: systematic random
   sampling, whole-section analysis, or selection by the operator. Report
   selection by an unblinded operator as a bias risk, and report field
   selection that is not described at all.

4. Blinding. Determine whether the person scoring or measuring was blinded to
   group. Report its absence, and note that this matters most for manual
   scoring and for any threshold set by eye.

5. Thresholding. Determine how positive signal was defined: a fixed intensity
   threshold, an automatic method named explicitly, or a threshold adjusted per
   image. Report thresholds adjusted per image without a stated rule, and
   report any thresholding described only as done in software.

6. Acquisition consistency. Determine whether exposure, gain, laser power and
   magnification were held constant across compared groups. Report where this
   is unstated, since differing acquisition settings invalidate intensity
   comparison.

7. Intensity claims. Where fluorescence intensity is compared across groups,
   determine whether the comparison is quantitative or qualitative, and whether
   background subtraction is described.

8. Scale bars and labels. Report panels without a scale bar and legends that do
   not state magnification. Where you can see the figure, report scale bars
   present in the image but not defined in the legend.

Report each issue with the exact panel and the specific addition needed.

The brief is ordered deliberately, and the order is the finding. Item 1 is first because a missing quantification cannot be repaired by improving items 2 through 8, and items 2 to 5 sit above 6 to 8 because a sampling or blinding defect changes the number while an undefined scale bar only changes what a reader can check. If you are writing your own version, see how to write a custom reviewer agent and the custom agent slots that hold it.

What this agent catches

Quantification versus illustration: why item 1 is first

A representative image is an existence proof, not a measurement, and this agent reports every between-group claim that rests on one. A micrograph shows that the staining pattern the authors describe can occur in that condition. It cannot show that the pattern occurs more often, more brightly or over a larger area in one group than another, because a single panel per group has no denominator and no variance.

The tell in the manuscript is a results sentence with a comparative verb and a figure callout that points only at panels. “Treatment reduced Ki-67 staining (Fig. 3B, C)” where 3B and 3C are two micrographs is an unquantified comparison, however obvious the difference looks. The reviewer’s question is not whether the images differ. It is how many images were looked at before those two were chosen.

Where the manuscript does carry a quantification, the agent checks that the quantification supports the sentence rather than a neighbouring one. Area fraction does not support a claim about cell number; mean intensity per field does not support a claim about the proportion of positive cells; a positivity index over whole sections does not support a claim about staining confined to the invasive margin. Mismatch between what was measured and what is asserted is the same defect the overclaim check reports in text, arriving through a figure instead.

The remedy is stated in one sentence and it is rarely convenient: quantify every panel that carries a comparative claim, on every field you acquired, or convert the claim to a description of what the image shows.

Common questions about representative images and quantification

What does an AI reviewer actually check for in image quantification?

An AI reviewer briefed for image quantification checks eight things in order: whether a comparative claim has a quantification behind it at all, sampling depth, field selection, blinding, thresholding, acquisition consistency, background subtraction, and scale bars with magnification. The first item is deliberately first, because a missing quantification cannot be fixed by improving any of the other seven.

Is it OK to just show representative images without quantifying them?

Only when the surrounding text makes no comparative claim. A representative image supports a descriptive statement — “staining appeared cytoplasmic in this condition” — but not a comparison between groups. If a results sentence uses a comparative verb (“increased”, “reduced”, “more”) and the figure callout points only at panels, that sentence needs a quantification, not a better image.

My reviewer wrote “please provide quantification across all animals” — what exactly do they want?

A number derived from every animal in the experiment, not the two or three whose images appear in the figure. Report how many fields per section and sections per animal fed into that number, report the summary statistic per animal rather than pooling fields across animals as if they were independent, and give the test used to compare groups.

The sampling hierarchy behind a histology number

Every histology quantification sits on a three-level hierarchy — fields within sections, sections within animals, animals within groups — and a figure legend that names only one of the three is incomplete. “n = 5 fields” leaves the reader unable to tell whether five fields came from five animals or from one. “n = 3 mice” leaves the reader unable to tell whether the estimate for each mouse came from one field or forty.

Report all three. The convention that survives review reads: “Five randomly selected 20× fields were scored per section, three non-adjacent sections per animal, and six animals per group; each point is the per-animal mean and n = 6.” That sentence fixes items 2 and 3 of the brief at once, and it makes the unit of analysis unambiguous.

Where the sampling effort should go is a solved problem in stereology. Gundersen and Østerby set out the answer in “Optimizing sampling efficiency of stereological studies in biology: or ‘do more less well!’” (Journal of Microscopy 1981;121(Pt 1):65–73): because between-animal variance usually dominates the total, precision is bought far more cheaply by adding animals than by adding fields within an animal. Twenty fields from three mice is a worse experiment than five fields from twelve mice, at similar bench time.

That guidance holds only when between-animal variance genuinely dominates for the assay in hand, and that is an empirical question rather than a given. A stain with strong focal or regional heterogeneity within an animal — a lesion confined to one lobe, a gradient radiating from an injection site — can invert the ratio, in which case the precision is bought within the animal, not across animals. A nested variance-components analysis (a mixed model or nested ANOVA with field nested in section nested in animal) run on a small pilot dataset settles the question for a specific assay, rather than importing Gundersen and Østerby’s general finding unchanged.

Counting objects from a two-dimensional section has its own bias, because larger objects are more likely to be intersected by the section plane. The disector — Sterio’s design-based counting method (Journal of Microscopy 1984;134(Pt 2):127–136) — removes that size dependence by counting objects that appear in one section of a pair and not the other. If your manuscript reports cell number per unit area from single sections, expect a reviewer with a stereology background to ask why.

The sampling hierarchy is also exactly the nesting that produces an inflated n, which is why this agent and the unit of analysis agent find each other’s leftovers. See what is pseudoreplication for the statistical consequence, and reviewer says n is not independent for the comment it produces.

How fields should be chosen

Field selection is the item that decides whether the number means anything, and there are exactly three defensible answers: whole-section analysis, systematic uniform random sampling, or a pre-specified rule applied blind. Everything else is the operator choosing what to measure.

Whole-section analysis is the strongest and is now usually feasible: a slide scanner produces the entire section, and a segmentation macro applied to the whole tissue removes field selection from the study altogether. Systematic uniform random sampling places a grid over the section with a random start and takes every field the grid lands on, which gives uniform coverage without the operator’s hand. A pre-specified rule — “the five fields nearest the tumour margin, moving clockwise from the 12 o’clock position” — is defensible when the biology demands a region, provided the person applying it does not know the group.

Whole-section and grid-random sampling remove the operator’s hand from which field gets chosen, but neither removes bias from how the field is then measured: a segmentation macro whose parameters were tuned by comparing outputs against the expected group difference reintroduces the same problem one step downstream. Fix the macro’s parameters on a training set that excludes the images under comparison, then apply it unchanged. A pre-specified region rule is also not automatically the weaker option — where the biology is genuinely regional, a whole-section average can dilute a real effect confined to the margin into noise, so the disqualifying feature is never that a region was chosen, only that the choice was made by someone who had already seen which group produced it.

The failure mode has a recognisable signature in the methods: “representative fields were selected for quantification”. That phrase describes a selection made by a person who has already seen the staining, and it means the quantification inherits whatever the representative image was chosen to show. It is the imaging form of the problem described in p-hacking — the analysis choice is made after the data are visible, and there is no record of the choices not taken.

State the macro. If fields were segmented in ImageJ or Fiji, the reviewer needs the ordered operations — channel, background subtraction and its radius, threshold method, size and circularity filters for Analyze Particles, and any watershed step — not the sentence “images were analysed in ImageJ”. A macro deposited alongside the paper answers the whole of items 3 and 5 at once, and it fits the same repository deposit as your data availability statement.

Questions about sample size and field selection

Is n=3 mice enough for an IHC quantification?

It depends on where the variance sits, not on a fixed rule. Gundersen and Østerby’s stereology work found between-animal variance usually dominates, which is why three mice sampled heavily can be a weaker design than six mice sampled lightly. A reviewer asking this wants the power calculation or the pilot variance behind the number, not a larger number offered without one.

How many fields per section do I actually need to count?

There is no fixed count that applies across stains and tissues, because the answer depends on how heterogeneous the staining is within a section — something you cannot know until you have measured it once. Report the number you used, how it was chosen, and whether whole-section or systematic random sampling was available instead of picking fields by eye.

Reviewer 2 says my sample size is too small for the histology — what do I say back?

Show the calculation behind the number rather than defending the number itself. Report the variance components from pilot data (or the closest comparable published dataset) and the effect size the study could detect; if the study was not powered in advance, say so plainly and report the achieved power for the effect actually observed.

Blinding the person who scores

Blinding matters most where a human eye assigns the number, and its absence is among the most common omissions in this area. The ARRIVE guidelines 2.0 make it Essential 10 item 5, worded: “Describe who was aware of the group allocation at the different stages of the experiment (during the allocation, the conduct of the experiment, the outcome assessment, and the data analysis)” (Percie du Sert et al., PLOS Biology 2020;18(7):e3000410). Blinding earns a place in the Essential 10 precisely because surveys of the animal literature keep finding it reported in a minority of papers.

The measured consequence is large. Bebarta, Luyten and Heard reviewed 290 animal studies with two or more groups presented at Society for Academic Emergency Medicine meetings between 1997 and 2001, and found that studies lacking blinding, and studies lacking randomisation, were significantly more likely to report a positive result — odds ratios of roughly three in each case, and higher again for studies using neither safeguard (Academic Emergency Medicine 2003;10(6):684–687). That is an effect on reported findings comparable in size to many of the treatment effects being reported.

Blinding image scoring is cheap in a way blinding an in-life experiment is not. Export the acquired images to a folder, rename them with a script to random identifiers, keep the key in a separate file, score the renamed set in one sitting, then rejoin the key. Two independent scorers with a stated agreement statistic — a weighted kappa for ordinal scores, an intraclass correlation for continuous ones — is stronger still, and takes an afternoon.

Blinding has a real limit worth naming rather than glossing over: a sufficiently obvious phenotype — a visibly larger tumour, an albino versus a pigmented strain, gross morphological differences in a knockout — can be identified from the image regardless of how the file is named, and ARRIVE 2.0 item 5 asks authors to describe this rather than assume blinding was achieved. Where the phenotype itself is unblindable, the honest response is to say so, quantify with an automated, fixed-parameter pipeline instead of a human score, and report any manual step as a stated limitation rather than a completed safeguard.

Note the distinction the brief draws. Automated segmentation with a fixed threshold applied identically to every image is blind by construction and needs no extra step. A threshold nudged per image by a person who knows the group is manual scoring wearing software’s clothes. Randomisation, blinding and exclusions across the whole experiment belong to the in vivo rigour agent; this agent covers only the measurement step.

How a threshold becomes the result

A threshold defines what counts as positive, so choosing it while looking at the outcome makes positivity a function of the result. The agent reports three distinct states: a fixed threshold stated numerically, a named automatic method, and a threshold adjusted per image without a stated rule. Only the third is a finding on its own, but the second is only adequate when the method is named.

“Thresholded in ImageJ” does not name a method. The Auto Threshold plugin implements seventeen algorithms — Default, Huang, Huang2, Intermodes, IsoData, Li, MaxEntropy, Mean, MinError(I), Minimum, Moments, Otsu, Percentile, RenyiEntropy, Shanbhag, Triangle and Yen — plus a Try all option that renders every one of them side by side. Otsu and Li can differ substantially on the same dim immunofluorescence field, and running Try all and picking the panel that separates your groups best is threshold selection by outcome. Name the method, apply it to every image in the comparison, and say whether it was computed per image or once on a pooled histogram.

Naming the algorithm is necessary but not sufficient, because Otsu and Li both assume the pixel-intensity histogram is close to bimodal, and a dim or sparsely positive immunofluorescence image frequently is not. On a histogram with one dominant background peak and a long, low tail of positive pixels, Otsu tends to place the cut well inside the background and inflate the positive fraction, while Li’s entropy-based criterion can behave differently again on the same distribution. Check the histogram the threshold came from, not only which algorithm’s name goes in the methods.

The threshold also depends on the implementation and not only on the algorithm, which is why two people who both ran “Otsu” on the same image can report different numbers with neither of them making a mistake. Otsu’s criterion selects a histogram bin, and it does not fix whether the reported threshold is that bin’s lower bound, its upper bound or its midpoint, so independent implementations resolve the question differently. ImageJ’s own Auto Threshold documentation records two further sources of divergence inside Fiji alone: the Image › Adjust › Threshold… applet cannot be told to ignore the extremes of the histogram while the Auto Threshold plugin can, so the same named method reached by the two menu paths can return different results; and on a 16-bit image the applet first compresses the histogram into 256 bins while the plugin works across the full greyscale range. For quantitative work, name the software and its version alongside the algorithm, and put every image in a comparison through the same path.

Where scoring is ordinal rather than pixel-based, use a published scale with fixed cut-offs rather than an in-house scheme. The H-score is defined as (% of cells at weak intensity × 1) + (% at moderate × 2) + (% at strong × 3), giving a range of 0 to 300. The Allred score sums a proportion score of 0 to 5 and an intensity score of 0 to 3 for a total of 0 to 8, with 0 to 2 read as negative and 3 to 8 as positive. Both are ordinal, both are scored by eye, and both therefore require item 4 — blinding — to mean anything.

The reviewer question that follows is whether the conclusion survives a different threshold. Reporting the primary result plus the same analysis under one alternative method is a two-line sensitivity check that pre-empts it, and it reads far better than the limitations paragraph you would otherwise write.

Questions about thresholding and blinded scoring

Why do Otsu and Li give me different thresholds on the same image?

Because the two methods optimise different statistics of the same intensity histogram, and dim or sparse immunofluorescence images often lack the clean bimodal distribution both assume. Otsu minimises within-class variance; Li maximises cross-entropy between foreground and background. Naming which one you used is necessary but not sufficient — check that its bimodal assumption actually fits your histogram before trusting its output.

I ran Otsu in ImageJ and my collaborator ran Otsu in Python — why don’t we get the same threshold?

Because “Otsu” names a criterion, not a single implementation. The criterion picks a histogram bin and does not fix whether the threshold reported is that bin’s lower bound, upper bound or midpoint, so different libraries resolve it differently on identical pixels. Fiji disagrees with itself for the same reason: the Image › Adjust › Threshold… applet compresses a 16-bit histogram into 256 bins and cannot be told to ignore the histogram extremes, while the Auto Threshold plugin does neither. Neither of you is wrong. Report the software, its version and the menu path alongside the algorithm, and run every image in the comparison through one of them.

How do I blind myself when I’m the only one scoring my own slides?

Rename the image files to random identifiers with a script before looking at any of them, keep the group key in a separate file you do not open until scoring is finished, and score the entire set in one sitting rather than in group-labelled batches. This is the cheapest fix in the whole review and it takes under an hour of setup.

Will an AI reviewer flag how I set my positive threshold?

Yes — it distinguishes a fixed threshold stated numerically, a named automatic method, and a threshold adjusted per image with no stated rule, and only the last of the three is a finding by itself. A named method applied identically across every image in the comparison, with the algorithm stated, passes this check.

Acquisition settings and why intensity comparison fails without them

Fluorescence intensity is comparable across groups only when the images were acquired identically, so exposure time, gain or detector voltage, laser power, pinhole, objective and numerical aperture, bit depth and zoom all belong in the methods as numbers. A control image taken at 200 ms and a treated image at 400 ms differ in brightness for a reason that has nothing to do with biology, and no amount of downstream normalisation recovers the comparison.

Saturation is the failure the eye does not catch. An 8-bit image clips at pixel value 255, a 12-bit camera at 4,095 and a 16-bit camera at 65,535, and every photon arriving after the clipping point is discarded. A bright group imaged at settings tuned for the dim group will show saturated pixels, and the measured difference will be smaller than the true one — a defect that pushes toward the null and so passes unchallenged when the result is positive. Check the histogram, not the display.

Sampling density is the other acquisition setting reviewers ask about in quantitative work. The Nyquist criterion requires a sampling frequency greater than twice the highest spatial frequency the optics transmit, which in practice means a pixel spacing well below the resolution limit rather than equal to it — and the required interval falls as wavelength shortens and numerical aperture rises, so it has to be computed for the objective and channel actually used rather than assumed from a previous experiment. Scientific Volume Imaging publishes a Nyquist calculator that returns the lateral and axial intervals for a given objective, medium and wavelength. Undersampled images cannot support measurements of object size or count, whatever the segmentation does afterwards.

Nyquist is a floor, not a target to exceed by as much as possible. Sampling well beyond the criterion buys no additional resolvable information while multiplying acquisition time, file size, and — for live or repeatedly imaged fluorescent samples — photobleaching and phototoxicity from the extra light dose collected. Match pixel spacing to the criterion for the objective and channel actually in use; oversampling by an order of magnitude is a different design flaw from undersampling, not a safer default.

Display adjustments are governed by a rule worth quoting exactly. The Journal of Cell Biology instruction, set out by Rossner and Yamada in “What’s in a picture? The temptation of image manipulation” (Journal of Cell Biology 2004;166(1):11–15), is that “No specific feature within an image may be enhanced, obscured, moved, removed, or introduced”, and that “Nonlinear adjustments (e.g., changes to gamma settings) must be disclosed in the figure legend”. Apply linear adjustments equally across the whole image and equally across compared panels, disclose anything nonlinear, and quantify from the raw file rather than the display copy.

Background subtraction and colocalisation claims

Background subtraction changes the number, so the method and its parameter belong in the methods section alongside the threshold. ImageJ’s Subtract Background uses the rolling-ball algorithm of Stanley Sternberg (“Biomedical Image Processing”, IEEE Computer, January 1983). ImageJ’s documented rule of thumb is that for 8-bit or RGB images the radius should be at least as large as the radius of the largest object in the image that is not part of the background. The same documentation gives a different rule for 16-bit and 32-bit images, scaling the radius inversely with the pixel value range and suggesting values around 0.2 to 5 for a 16-bit image — guidance that image analysts on the ImageJ community forum have tested and disputed, on the reasoning that a rolling ball is a structuring element measured in pixels and so should track object size rather than intensity range, and on the observation that a radius below one pixel changes the image not at all. Size the radius from your largest foreground object whatever the bit depth, preview the subtracted result against the original rather than trusting either rule unseen, and apply one radius across every image in the comparison. Either way the radius is a number, and it belongs in the methods. A radius smaller than the objects being measured erodes their signal; the same radius applied to a dim group and a bright group removes different amounts. Report the radius.

The alternative is an explicit background region: measure mean intensity in a defined region of each image containing no cells and subtract it from each object’s mean, stating how the region was chosen. Either approach is defensible. Neither is defensible unstated.

Colocalisation claims carry a specific and frequently mishandled requirement. Pearson’s correlation coefficient measures whether two channels’ intensities co-vary; Manders’ coefficients M1 and M2 measure what fraction of each channel’s signal overlaps the other. A sentence claiming “protein A colocalises with protein B” is an overlap claim, and Pearson’s alone does not test it. Dunn, Kamocka and McDonald set out the distinction and the practical pitfalls in “A practical guide to evaluating colocalization in biological microscopy” (American Journal of Physiology — Cell Physiology 2011;300(4):C723–C742). Where thresholds are needed for Manders’ coefficients, the automatic method of Costes and colleagues (Biophysical Journal 2004;86(6):3993–4003) removes the operator from that choice and should be named when used.

Report the coefficient, the threshold method, the number of cells and the number of independent experiments, and state whether a randomisation or scrambled-channel control was run. Without that control, a high coefficient in a densely stained compartment says little.

Two further conditions are worth checking before trusting either coefficient. Pearson’s r has a failure mode specific to multi-channel fluorescence: spectral bleed-through between channels manufactures a positive correlation unrelated to the two proteins’ real spatial relationship, so a colocalisation claim needs a single-stain control confirming each channel is captured without significant crosstalk before the coefficient means anything. Costes’ automatic thresholding is not assumption-free either — it estimates the threshold from the point where the two channels’ correlation becomes indistinguishable from a randomised control, which requires enough non-overlapping area in the image to fit that null; a field that is densely and near-uniformly positive in both channels can leave the method too little unbound area to work with, and the fit should be checked rather than trusted by default.

Scale bars, magnification and what a legend must carry

A scale bar is the only feature of a micrograph that lets a reader check a size claim, and a bar drawn in the panel but not defined in the legend is not a scale bar. The agent reports both states: panels with no bar at all, and bars whose length is never stated.

A figure legend for a quantified micrograph should carry, in one place: the scale bar length; the objective magnification and numerical aperture; the stain or antibody with its target; the number of fields, sections and animals behind each point; what each point represents (a field, an animal, an independent experiment); the summary statistic and the dispersion measure; and the test applied. Legends that give a mean and an unlabelled error bar force the reader to guess between standard deviation, standard error and a confidence interval, which changes how the figure reads by a factor of √n.

Magnification stated as “×20” without the objective’s numerical aperture is incomplete for quantitative work, because two 20× objectives with different NA resolve differently and collect different amounts of light. State both. Where panels are digital crops of a larger field, say so, and give the scale bar for the crop rather than the parent image.

Frequently asked questions about background subtraction, vision and colocalization

How do I choose the rolling ball radius for background subtraction?

Size it from the largest foreground object you need to keep, so the ball rolls under the background but not under your cells: the documented rule of thumb is a radius at least as large as that object’s radius. Too small and it eats into the signal you are measuring; too large and it stops correcting an uneven background. Preview the result against the original before committing to a value, use the same radius for every image in the comparison, and state the number in the methods — a radius tuned image by image is the same defect as a threshold tuned image by image.

Does an AI reviewer need to see my figures, or just the text?

It works from the methods and legends alone, but checking a scale bar or the content of a panel needs vision enabled — a caption cannot substitute for the image when the question is whether a bar is drawn but never defined. Enable vision and the capability figures_tables for this agent specifically.

Can an AI reviewer detect image manipulation or duplicated panels?

No. This agent reviews whether a described measurement supports the claim made about it; it does not compare pixel blocks across panels looking for duplication or splicing. That is a separate discipline carried out by trained human image forensics, distinct from what this agent is briefed to check.

My colocalization reviewer asked for Manders’ coefficients — I already reported Pearson’s r, isn’t that the same thing?

No, and this is a common mix-up. Pearson’s r measures whether two channels’ intensities co-vary; Manders’ M1 and M2 measure what fraction of each channel’s signal physically overlaps the other. A claim that two proteins colocalise is an overlap claim, and Pearson’s alone does not test it — report both, with the threshold method named.

How to detect these problems in your own figures before a reviewer does

Image quantification defects are detectable in your own draft, and the way to find them is to read your figures the way this agent reads them: claim first, then number, then method. Four passes catch most of it.

Pass one — highlight every comparative verb in the results. “Increased”, “reduced”, “enhanced”, “more”, “fewer”. For each one, follow the figure callout and confirm a quantification exists with a test attached. Sentences whose only support is a panel are your item 1 findings, and they are usually a small, fixable set.

Pass two — write the three numbers next to every micrograph panel. Fields, sections, animals. Any panel where you cannot write all three from your own records is a panel where the legend cannot state them either. This is also where the unit-of-analysis problem surfaces, because writing “40 fields, 1 section, 2 mice” next to an n of 40 makes the mismatch visible in a way the figure never does.

Pass three — open the raw files and check the metadata. Exposure, gain, laser power and objective are recorded in the acquisition file. Compare them across groups. Where they differ, either re-acquire or downgrade the intensity claim to a qualitative one; there is no third option.

Pass four — reconstruct the analysis from the methods alone. Hand the methods paragraph to a colleague who did not do the experiment and ask them to state the threshold method, the background subtraction radius and the field selection rule. Whatever they cannot recover is what a reviewer will ask for, and it is the same methods objection whether it arrives from a colleague or from a referee.

What reviewers say when image quantification is mishandled

Reviewers rarely write “image quantification integrity”. They write the specific version, and these comments hold papers for months.

“Figure 3 shows representative images only; please provide quantification across all animals with the statistical test used.” “The legend states n = 12 without specifying whether this refers to fields, sections or animals.” “How were fields selected for analysis, and was the operator blinded to group?” “The methods state that images were thresholded in ImageJ; please name the thresholding algorithm and state whether it was applied per image.” “Were acquisition settings held constant between the control and treated groups? Please report exposure, gain and laser power.” “The intensity comparison in Figure 5 is not interpretable without evidence that the signal was within the linear range of the detector.” “Scale bars are present in panels A–D but their length is not given.” “Colocalisation is claimed on the basis of Pearson’s coefficient; please report Manders’ coefficients and state the thresholding method.” “Please confirm that no nonlinear adjustments were applied, or disclose them in the legend.”

The last of those is the one authors handle worst, because the honest answer is often that a gamma adjustment was made in a figure-preparation program and never recorded. Disclosure in the legend closes it. Denial that the raw files contradict does not, and it converts a methods query into an integrity query. When these comments arrive together, answer them as a set rather than one by one — see how to write a response to reviewers.

What to do when the obvious remedy is unavailable

The obvious remedy — re-acquire the images under matched settings, with a blinded scorer and a pre-specified field rule — is often out of reach, and the useful question is what to do then. Three situations recur.

The animals are gone and the blocks are exhausted. Re-score what you have, blind. Export the existing images, randomise the filenames, score the whole set in one sitting, and report the blinded re-score as the primary analysis with the original as a sensitivity check. This recovers item 4 without a single new section, and it is the highest-value hour available to most manuscripts at revision.

Acquisition settings differed between groups. Intensity comparison is not recoverable, and no normalisation makes it so. What is recoverable is a settings-independent measure: the fraction of cells positive at a fixed threshold set on a shared control, the number of objects per unit area, or a ratio between two channels acquired in the same session. Report the new measure as the claim and describe the intensity difference qualitatively, or state plainly that the comparison is not quantitative.

Only representative panels exist because quantification was never planned. Quantify retrospectively on every field you acquired, not on the panels you chose, and say in the methods that all acquired fields were analysed. Where too few fields were acquired to support a test, the honest move is to narrow the claim to what the images show and remove the comparative verb — a far smaller loss than defending an unsupported comparison through two rounds of review.

In all three cases, name the limitation, its direction and its likely magnitude. “Scoring was not blinded, which in comparable animal work is associated with an odds ratio of roughly 3 for reporting a positive result” is a specification. “Results should be interpreted with caution” is not.

How this differs from the built-in review

The image quantification agent differs from PerfectPaper’s standing figure specialists by reviewing laboratory practice rather than figure design: the standing team’s two figure agents — figure integrity and data-ink, and figure-to-text-to-caption consistency — read figures carefully but are not briefed on the sampling hierarchy behind a histology quantification, on how a threshold should be set, or on why acquisition settings determine whether an intensity comparison means anything. Those are laboratory practice questions rather than figure design questions.

Note the overlap with pseudoreplication: fields per animal is exactly the nesting that produces an inflated n. Running both is a reasonable use of two slots. Full set: AI peer review for preclinical cancer research.

The agent is also not a forensic tool. It reviews the description of a measurement against the claim the measurement supports; it does not compare pixel blocks between panels looking for duplication, and it should not be described as doing so. Reporting compliance against a named checklist is a separate job again, covered by the ARRIVE guidelines page.

Review my manuscript

Related agents

The four agents below share a review with this one, and three run together. See AI peer review for preclinical cancer research for the full set of five preclinical agents.

• Unit of analysis and pseudoreplication — use it when a figure legend counts cells, tumours or wells rather than animals.
• In vivo experimental rigour — use it for randomisation, blinding, exclusions and the tumour volume formula.
• Flow cytometry reporting — use it when a percentage has no named parent population.
• Model-to-human relevance — use it when an immune or microenvironment claim may exceed the host used.