Diagnostic model, 2,140 images

Machine learning research poster

A model that detects diabetic retinopathy from smartphone fundus images, laid out as an A0 landscape board. It is built around the figure, which is what a machine learning poster usually needs, and the text does the two things these posters most often skip: it describes the split honestly, and it says plainly what the result does not show.

Create a poster with OneCraftA0 Landscape, printed at 1189 by 841 millimetres

The whole board

The poster at full size, exactly as it prints. Every number, citation and caption on it was written for this example, so the layout is being judged on real content.

Machine learning detection of diabetic retinopathy from smartphone fundus images in low resource clinics

J. Chen ¹, M. Garcia ¹,², T. Adebayo ²

1 Department of Computer Science, Northgate University · 2 Institute for Tropical Medicine, New Kent

Abstract

Referral for diabetic retinopathy depends on a trained grader and a tabletop fundus camera, neither of which is present in most district clinics. Patients are therefore referred on symptoms, by which point the disease that is easiest to treat has usually been missed. We trained a convolutional model on 2,140 images captured with a $60 smartphone lens adapter and graded by two ophthalmologists, and tested it on a held out set of 428 images from three clinics that contributed no training data at all. Sensitivity for referable disease was 91.4% (95% CI 86.2 to 95.1) and specificity 88.7% (95% CI 84.0 to 92.4), measured against a two grader reference standard with a third grader adjudicating disagreements. Disagreement between the two human graders on the same images was 7.9%, so the residual error of the model sits inside the band where the humans disagree with each other. The failure cases were dominated by media opacity, mostly cataract, rather than by disease grade, which points the next round of work at image quality at the point of capture rather than at the model.

METHODS

Data. 2,140 images from 1,070 patients, both eyes, captured on a smartphone with a clip-on 20 dioptre lens.

Reference. Two ophthalmologists graded independently on the international scale; a third adjudicated disagreements.

Model. EfficientNet-B0 pretrained on ImageNet, fine tuned for 40 epochs, images resized to 512 by 512.

Split. Patient level split. The test set came from three clinics that contributed nothing to training.

Analysis. Sensitivity and specificity for referable disease, with bootstrap confidence intervals over 2,000 resamples.

91.4%

sensitivity for referable disease

95% CI 86.2 to 95.1, n=428

Sensitivity 91.4%, specificity 88.7% on unseen clinics

Runs on a $60 lens adapter and an existing phone

Errors driven by media opacity, not disease grade

Quality filter lifts sensitivity to 94.8%

Sensitivity

Specificity

Figure 1. Sensitivity and specificity by disease grade on the held out test set.

Results

On 428 held out images the model reached 91.4% sensitivity (95% CI 86.2 to 95.1) and 88.7% specificity (95% CI 84.0 to 92.4) for referable disease. Performance was weakest on mild disease, where sensitivity fell to 78.6%, and strongest on severe and proliferative grades, where it exceeded 95%. That pattern matters less than it first looks. Mild disease is not referable under the local pathway, so the misses that carry clinical cost are concentrated in exactly the grades the model handles best, and the mild cases it misses are returned to routine annual screening rather than lost. The test set held 171 referable eyes and 257 that were not referable, so 91.4% sensitivity means 15 misses. Of those 15, nine came from images the graders had already flagged for media opacity, mostly cataract, and three from images where the flash reflected off the lens adapter. Only three were failures on a clear image. Excluding images that failed an automated quality check lifted sensitivity to 94.8% while discarding 11.2% of the set. That is the trade a clinic has to weigh: a better model on fewer images, against the cost of asking one patient in nine to come back. In a district where the return journey is an hour each way, the arithmetic may favour accepting the lower figure. Inter grader disagreement on the same 428 images was 7.9%. The model is therefore wrong about as often as the two ophthalmologists were wrong about each other, which sets a practical ceiling on what any further tuning against this reference standard can achieve. Throughput was 1.8 seconds per image on a mid range phone with no network connection, so the pathway does not depend on a clinic having reliable data, and no image leaves the device. Two things this study does not show are worth stating plainly. The first is anything about outcomes. We measured agreement with a grading standard, not whether a single patient was treated earlier or kept their sight, and the step from one to the other runs through a referral pathway we did not test. The second is generalisation beyond this population. All 1,070 patients were recruited in one region, cataract prevalence here is high, and a model whose failures are dominated by media opacity will behave differently where that is not true. Cost is the argument for the approach rather than accuracy. The lens adapter is $60 against roughly $18,000 for the tabletop camera the district hospital uses, the model runs on phones the clinics already own, and a screening visit needs a health worker with two hours of training rather than an ophthalmologist. At the observed sensitivity that trade buys a great deal of coverage for a modest loss of accuracy, and coverage is the binding constraint in this district.

Block by block

What each block on the board is for, in the order a reader walks it.

Title band and authors: Title, generated author superscripts, both affiliations and a QR to the code repository, on a single band across the top of the landscape board.
Abstract: The task, the dataset, the headline metrics with intervals, and the fact that inter grader disagreement was 7.9%, which sets the practical ceiling on the result.
Methods as a labelled protocol: Five entries: data, reference standard, model, split and analysis. The split entry is the one reviewers stop at, so it says explicitly that the test clinics contributed no training data.
The dominant figure: Sensitivity and specificity by disease grade across the full width of the board, which is the figure a reader will look at before reading anything.
Key number: One stat block carrying 91.4% sensitivity with its interval and its denominator, rather than a row of metrics competing for attention.
Results: Four paragraphs: the headline metrics, the error breakdown, what the quality filter trades, and the human disagreement ceiling. The weakest grade is reported rather than buried.
What it does not show: A paragraph on the two limits that matter: no outcome was measured, and the failure profile is specific to a population with high cataract prevalence.
Cost and takeaways: The argument for the approach is cost rather than accuracy, stated with the numbers, followed by four checked takeaway points. The three references are invented for this fictional study, because the reference block is required on this layout; on your board, replace them with your sources.
How to adapt this board: Describe the split before anything else: how many cases, from where, and what the model never saw. On a diagnostic board that sentence decides whether the rest is worth reading. Then pick the one curve that is the result and give it the hero figure, because a board with four curves has no result. Errors go by class in the results text, since an overall accuracy hides the failure that matters clinically. The key number block takes the single headline metric with its interval, and the repository link goes behind the QR code where the hyperparameters and the code can live in full.

What makes this board work

The split is described, not assumed

The methods block states that the split was at patient level and that the test set came from three clinics contributing nothing to training. Most model posters say "held out set" and leave the reader to guess whether the same patient appears on both sides.

Errors are explained, not just counted

Fifteen misses are broken down into nine with media opacity, three with flash reflection and three genuine failures on a clear image. That turns a number into a direction for the next piece of work.

It states what it does not show

A paragraph says outright that no outcome was measured and that the population is one region with high cataract prevalence. A poster that names its own limits is more persuasive than one that does not, because reviewers ask anyway.

Questions people ask

What goes on a machine learning research poster?

The task and why it matters, the dataset with its size and how it was labelled, the split, the model, the metrics with confidence intervals, an error analysis, and the limits. A confusion matrix or a curve usually carries the result better than a table.

Should I show accuracy on a poster?

Rarely on its own. For a screening task, sensitivity and specificity with intervals say what accuracy hides, because a model can score well on accuracy by rarely predicting the rare class.

Landscape or portrait for a machine learning poster?

Landscape suits a single dominant figure and a wide reading order, which is the usual shape for a model result. Portrait suits a text heavier study. Check what the conference specifies first.

Can I put my own figure on it instead of a chart?

Not on this layout. The hero figure is a chart block, so it takes one of ten chart kinds with your own values rather than an image file. Four layouts carry image figures, but those slots cannot be filled yet either, so plan on charts and paste your numbers.

Which chart kind suits a ROC or PR curve?

Line, with one series per model and the operating point named in the caption. The chart block accepts ten kinds, and line is the only one that keeps a curve honest. Never put a curve in an area chart, which fills under it and hides the comparison.

Should I list hyperparameters?

The ones that matter for reproduction, in one methods entry: learning rate, batch size, epochs, seed. Everything else goes in the repository the QR code links to. A board that lists forty settings is a board nobody reads.

Where does the data statement go?

In discussion, as a sentence that says what the model did not see: sites, scanners, years or populations that were not in the training data. On this board that is the single site and the two scanner models.

Build your own in about a minute

The button below opens the generator with this use case already described. Change the wording to match your own, generate, then edit anything you like.

Make my machine learning research poster

Sources

Written and checked by the OneCraft team. Last checked 30 August 2026.