Skip to main content
MP
← Back to Projects
Explainable AI · Healthcare · Computer Vision2026

RETINASIGHT

“Explainable AI pipeline for Diabetic Retinopathy screening in rural India.”

Role

Lead Developer (Team OnFocus)

Stack
PythonPyTorchResNet-50ONNXFastAPIMATLAB EngineOpenCVReactVite
Status

Delivered

Overview

RetinaSight is an offline-first, explainable AI screening pipeline for Diabetic Retinopathy (DR) built for Smart India Hackathon 2026 (Problem Statement 26038).

Designed for rural Primary Health Centres (PHCs) in India where ophthalmologist availability is severely limited, RetinaSight grades fundus captures across 5 ICDR severity levels (0 to 4) and provides visual lesion explainability via Grad-CAM.

Powered by a persistent MATLAB Engine interoperating with an asynchronous FastAPI backend, the pipeline includes sub-second image quality gates, green-channel CLAHE contrast normalization, retinal vessel tree segmentation, and automated Ayushman Bharat referral slip generation.

Screenshots

Key Features

→
Sub-Second Quality & Authenticity Gate - Evaluates blur variance and chromatic R/B spectrum to reject blurred scans or non-retinal photos in 6ms.
→
Green-Channel CLAHE & Vessel Segmentation - Normalizes contrast across peak hemoglobin absorption wavelengths and isolates the retinal vascular tree.
→
5-Class ResNet-50 Classifier - Deep learning classification head fine-tuned on clinical datasets with class-weighted loss and ordinal regression.
→
FOV-Constrained Grad-CAM Explainability - Generates saliency heatmaps strictly masked within the retinal field of view to visually isolate microaneurysms and hemorrhages.
→
Ayushman Bharat Referral Generator - Generates standardized clinical referral slips and camp rosters aligned with NPCBVI guidelines.

Tech Deep Dive

The inference architecture integrates MATLAB Engine asynchronously into FastAPI using thread pool executors. Deep Learning Toolbox dlnetwork executes calibrated ResNet-50 weights, while Image Processing Toolbox performs morphological vascular segmentation. Visual saliency is generated natively with gradCAM() constrained to retinal aperture masks.

Next ProjectRAZORPAY ACP ADAPTER →