VESIT · BE Capstone 2026–27 · decision-support research prototype

RetinAI — ROP Screening

Deep-learning screening for Retinopathy of Prematurity from a single neonatal fundus photograph — with the audits, corrections and honest numbers that most demos leave out.

The interactive demo now runs locally. We deleted the hosted version of this app on 14 Aug 2026 to fully purge patient images that had survived in its git history — patient privacy outranks a convenient link, and we would make the same call again. The identical application runs on your machine in four commands below.

0.881device-controlled AUC (screening)

0.990sensitivity @ served threshold

0.69 ± 0.09macro-F1, ICROP staging 5-fold CV

8,024images · 1,760 infants · 4 cohorts

Run the app locally (~5 minutes, CPU only)

git clone https://github.com/Coldbari/RetinAI_BE_Project_2026_2027.git
cd RetinAI_BE_Project_2026_2027
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python scripts/get_weights.py     # ~180 MB, one time
python webapp/app.py              # → http://127.0.0.1:5002

You get the binary ROP screening result with a Grad-CAM heatmap and PDF report, plus a clearly-labelled 6-class ICROP staging research preview. Non-retinal images are rejected by a gradability gate; patient context is mandatory.

Project log book & code Model weights Project video (4½ min)

What makes this project different

Not the model — the measurement. Patient-grouped cross-validation with a held-out hospital and a still-unopened locked test set; a five-gate shortcut audit run before any number is trusted; a metadata-only baseline every model must beat; and a public record of the claims our own audits withdrew, including a single-fold "architecture win" that dissolved under full cross-validation. Every number in the log book traces to a committed artifact.

Not a medical device. Research prototype — no regulatory clearance, no prospective clinical validation. A qualified clinician confirms every finding.
Rutuja Bait · Pravar Rangnekar · Yash Shengale · Praharsh Nagpure — guide: Mrs. Amudha Kumar.