Explainable AI for Diabetic Retinopathy Screening in Rural India

The receipts 🧾

Every red flag on this statement, in full. These are the four places it bites.

1. Exhibit A

   Diabetic retinopathy classification is explicitly one of the most repeatedly submitted hackathon ideas and APTOS solutions are public, so the grading model on its own contributes nothing a judge has not seen

2. It gets worse

   The stated sensitivity and specificity targets are for referable DR on clean benchmark data, and reproducing them on the degraded portable-camera images the background describes is a materially harder problem the statement quietly conflates with the easy one

3. Still reading?

   IDRiD's pixel-level lesion annotations cover only a few dozen images, so microaneurysm and neovascularisation segmentation is being asked for on a data foundation far too thin to support the claim

4. And the finisher

   The Simulink capacity model is a stated deliverable that has nothing to do with the imaging and will be dropped by almost every team, which makes it both a risk and, if you actually build it, your clearest differentiator

The damage report 📉

Every score this statement earned, and what each one actually costs you.

• Feasibility

  3/5

  Buildable. Not comfortably. There is a week in here you have not planned for yet.

  All four datasets are public and linked and DR grading itself is very achievable, but the statement asks for much more than grading — sub-pixel microaneurysm detection, neovascularisation detection and lesion-level segmentation need the pixel-level IDRiD annotations which cover only a few dozen images, so the segmentation half rests on a far thinner data foundation than the classification half.

• Innovation scope

  2/5

  Nothing here is new. Your only edge is execution — and execution is also everyone else's only edge.

  Five numbered modules prescribe not just what to do but how — CLAHE and illumination normalisation for enhancement, the International Clinical DR scale for grading, Grad-CAM for explanation, Simulink for the workflow model — leaving very little methodological choice anywhere in the pipeline.

• Clarity

  5/5

  The ask is unambiguous, which quietly removes your favourite excuse.

  Among the most precisely specified statements on the portal: it names the enhancement techniques, every retinal structure to extract, the grading scale and its levels, quantified sensitivity and specificity targets, the explainability method and even the thirty-second clinician review budget.

• Acceptance potential

  2/5

  The numbers do not like you. Bring something the numbers cannot see.

  Diabetic retinopathy classification is one of the most cloned projects in existence and APTOS is a public Kaggle competition with thousands of published solutions, so your grading model is a fine-tune of something already benchmarked to exhaustion — the differentiators are the quality gate, the lesion-level explanation and the Simulink capacity model, all of which most teams will skip.

• Effort

  Massive

  A semester of work wearing a hackathon costume. Something is getting cut; decide what now, not in week five.

  Quality assessment, six distinct segmentation tasks, a graded classifier hitting stated clinical thresholds, a calibrated explainability layer with automated reporting, and a separate Simulink throughput model is at least three projects, and the lesion segmentation alone is a research effort.

• Demo-ability

  Easy

  Easy to demo — and so is everyone else's. Working is the floor here, not the achievement.

  A fundus image with a severity grade and a heatmap over the lesions that drove it is instantly comprehensible, and the quality-rejection behaviour gives you a second demo moment that most competing teams will not have.

The demo they will have already seen 🪞

Somewhere around 85–200 teams are heading here, and the description is doing the choosing for most of them. They will read the same brief, reach the same architecture, and build a version of the same demo you are planning. Being correct is the floor. If your five minutes could be swapped with the team before you and nobody in the room would notice, you have not picked badly — you have built predictably, which costs exactly the same and hurts more.

What survives 🛡️

The ground worth standing on when the questions start.

• Four benchmark datasets are linked including IDRiD, which is Indian and includes pixel-level lesion annotations, so you can genuinely evidence the explainability claim rather than only the classification
• The performance targets are stated numerically, so success is defined for you and you can report against published benchmarks instead of arguing about what good means
• The image quality gate is the most practically important and least attempted module — field fundus cameras produce ungradeable images constantly, and a system that knows when to refuse is a real clinical contribution