Explainable AI for Diabetic Retinopathy Screening in Rural India
MathWorks · Clean & Green Technology · Software
The grading half is a solved Kaggle problem you cannot win on, so if you take this, the quality-rejection gate and the lesion-level explanation have to be the entire pitch — build the Simulink model too, because it is the one deliverable nobody else will bother with.
Data: APTOS 2019 Blindness Detection: https://www.kaggle.com/c/aptos2019-blindness-detection ; IDRiD (Indian Diabetic Retinopathy Image Dataset): https://ieee-dataport.org/open-access/indian-diabetic-retinopathy-image-dataset-idrid ; DRIVE (vessel extraction): https://drive.grand-challenge.org/ ; Messidor-2: https://www.adcis.net/en/third-party/messidor2/
What it actually is
Diabetes damages the retina and eventually blinds people, and catching it early prevents almost all of that loss, but rural India has nowhere near enough eye specialists to screen everyone. Software can grade retinal photographs, except the existing tools give an answer without showing their working, which no doctor will accept. The ask is a screening pipeline that grades the image, shows the specific lesions behind its verdict, and refuses to grade photographs that are too poor to judge.
What to build
A five-stage MATLAB pipeline as the statement specifies: an image quality gate scoring focus, illumination and field of view, applying CLAHE and illumination normalisation to borderline images and rejecting ungradeable ones with recapture feedback; a structure segmentation stage localising the optic disc and fovea and extracting vessels, microaneurysms, exudates, haemorrhages and neovascularisation; severity grading on the International Clinical DR scale levels zero to four against the stated targets of over 90 percent sensitivity and 85 percent specificity for referable disease; an explainability layer producing Grad-CAM attention maps, lesion-level evidence tied to the clinical grading criteria and calibrated confidence, packaged into an annotated report an ophthalmologist can validate in under thirty seconds; and a Simulink model of the telemedicine pipeline sizing acquisition rate, bandwidth and review capacity for a district programme.
Smallest thing that wins the room
Feed a poor-quality fundus photograph and have the system reject it with recapture guidance, then feed a gradeable one and show the grade beside a lesion overlay where each flagged microaneurysm maps to the criterion that produced the level.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 72% of the 226 · #65 of 226 by expected field
A normal-sized field. Your idea has to be good, not miraculous.
Why: company-sponsored statements drew the smallest fields of all.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
2/5Diabetic 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.
Feasibility
3/5All 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
2/5Five 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
5/5Among 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.
Effort
MassiveQuality 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
EasyA 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.
In its favour
- Green flag: 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
- Green flag: 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
- Green flag: 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
- Green flag: The theme is filed under Clean and Green Technology, so teams browsing MedTech will not surface this at all
Against it
- Red flag: 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
- Red flag: 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
- Red flag: 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
- Red flag: 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
What you will be writing
- EfficientNet / ConvNeXt DR severity grading
- Grad-CAM and lesion-level attribution
- CLAHE and illumination normalisation preprocessing
- IDRiD pixel-level lesion segmentation
- temperature-scaled confidence calibration
- Simulink telemedicine throughput model
- Medical imaging
- Explainable AI
- Rural health screening
Prior art to read before you start
diabetic retinopathy severity grading · retinal lesion segmentation and localisation · explainable clinical decision support
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.