RetinaVision AI: Exposing the Ghost in the Black Box of Medical AI

Soumyajit Chaudhury — Wed, 28 Oct 2026 17:34:42 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine

What I Built

RetinaVision AI turns any rural Primary Health Centre (PHC) into a diabetic retinopathy (DR) screening point.

In rural India there is roughly one eye doctor for every 100,000 people. Diabetic retinopathy is silent in its early stages, so many patients reach a specialist only after irreversible vision loss, even though the blindness is preventable with timely screening.

A health worker uploads a retina (fundus) photo, even one taken on a smartphone. The system then:

Validates the image and rejects screenshots and random photos in about a second
Checks quality (sharpness, brightness, contrast, field of view) and enhances usable images
Detects lesions: segments the optic disc and vessels, and counts microaneurysms, hemorrhages and hard exudates
Grades the disease from Normal to Proliferative DR (Grade 0-4) with a confidence score
Explains itself with heatmaps showing what the model looked at
Refers high-risk patients to a specialist in one click, with a full PDF report attached

How I read the theme: every AI model has a ghost in the machine: a decision made somewhere inside layers of weights that nobody can see. In medicine, a ghost you can't question is dangerous. RetinaVision AI is built to make the ghost visible. Grad-CAM, Multi-Layer CAM, Occlusion and LIME let a doctor see where the model looked and why it reached its grade, instead of trusting a black box.

Demo

Code

Soumyajitttt / RetinaVisionAI

AI-assisted diabetic retinopathy screening for rural India, with Explainable AI. Demo Video Link ⇓

RetinaVision AI

AI-assisted diabetic retinopathy screening for rural India: simple, fast, and explainable.

RetinaVision AI turns any Primary Health Centre (PHC) into a diabetic retinopathy (DR) screening point. A health worker uploads a retina (fundus) photo, even one taken on a smartphone, and the system grades the disease, shows why it made that decision, and refers high-risk patients to a specialist in one click.

Demo video -> https://youtu.be/7Pvx6oEvqRg?si=dKT32cT2UV1CgszA

The Problem

In rural India there is roughly one eye doctor for every 100,000 people. It is not possible to detect symptoms early without medical assistance, so patients reach a specialist only after irreversible vision loss, even though this blindness is preventable with timely screening.

Our Solution

Screen at the PHC: health workers do first-line screening, so doctors are no longer the only line of screening.
AI grading: each scan is graded from Normal to Proliferative DR with a confidence score.
Explainable…

View on GitHub

Original project by Swastik; this is my fork.

How I Built It

React + Vite  ──HTTP/JSON──▶  Node/Express + MongoDB  ──matlab -batch──▶  Pipeline
 (frontend)   ◀─────────────       (backend)           ◀── result.json ──  (quality, segment,
                                        │                                   classify, explain)
                         Cloudinary (images) · Nodemailer (referrals)

Request flow: upload → JS pre-check → pipeline → images stored → result saved in MongoDB → shown in the UI.

Model: a CNN classifier trained for DR severity grading (Grade 0-4) [ADD: architecture, dataset, and accuracy/validation results]
Explainability suite: Grad-CAM, Multi-Layer CAM, a Cross-Method check (Grad-CAM vs Occlusion vs LIME), and layer Activations. Showing agreement or disagreement between methods stops a single heatmap from being over-trusted.
Image processing: CLAHE, illumination normalization and denoising to cope with low-quality field images
Backend: Node.js, Express, MongoDB, JWT auth, Socket.io, Multer, Sharp, Nodemailer
Frontend: React, Vite, Tailwind CSS
Workflow: separate clinician and doctor roles, a "My Patients" inbox, in-app case discussions, and a searchable referral history

Why Does Open Innovation Matter?

Explainability only works when you can open the model up. The explainability methods here (Grad-CAM, Occlusion, LIME) are all openly published techniques. They work because we own the model and can inspect its gradients and activations. A closed diagnostic API gives you a label and a score, with no way to ask it why.

The model is ours, not rented.

No per-scan API fees, which matters for under-funded PHCs screening at volume
Patient scans are analyzed on our own server rather than sent to a third-party inference API
We can retrain or fine-tune on local data, since fundus images from Indian smartphones and cameras look different from the clean clinical datasets most models are trained on
The grading pipeline can't be changed or shut down by a vendor

Honest limits: the pipeline currently runs on MATLAB, which is proprietary, and referral images are stored on Cloudinary. The next step is to port the pipeline to an open stack (PyTorch/ONNX) and self-host image storage, so that every layer, not just the model, is open and can run in a clinic with poor connectivity.

The Spirit of It

A black box can say "Grade 3" and be right, but a health worker or doctor has no reason to believe it. A heatmap glowing over the exact hemorrhages and exudates a specialist would look at is what turns "the AI said so" into "I can verify this." In a place where there's one specialist per 100,000 people, that trust decides who gets referred in time and who loses their sight.

Team

Built together with Swastik (@Swastik1901 on GitHub), who started the original RetinaVisionAI repository that this project builds on. He doesn't have a DEV account, so I'm crediting him here directly. Huge thanks to him for the foundation of this project.

RetinaVision AI is a decision-support and screening aid, not a replacement for a qualified ophthalmologist. All results should be reviewed by a medical professional before any clinical decision.

DEV Community: Soumyajit Chaudhury