Lightweight CNN for Retinal Fundus Images
A four-class retinal fundus image classifier reaching 91.14% precision with a smaller compute footprint than standard models — built so screening support does not require GPU-class hardware.
Key result
91.14% precision— Across 4 retinal classes
01Problem
Retinal screening depends on specialist review of fundus images, which is scarcest where the disease burden is highest. Published deep models reach strong accuracy but assume compute that low-resource clinics do not have.
02Objective
Design a compact CNN that classifies four retinal conditions accurately enough to be useful for screening triage, while staying small enough to run without specialist hardware.
03Architecture
- Fundus Images
- Preprocessing
- Lightweight CNN
- 4-Class Prediction
04Technology
- Python
- TensorFlow
- CNN
- OpenCV
05Implementation
Key technical decisions
- 01Designed a compact architecture with a constrained compute budget instead of fine-tuning a large pretrained backbone.
- 02Targeted precision as the headline metric, since a false positive sends a patient for unnecessary specialist review.
- 03Kept preprocessing identical across splits so evaluation reflects deployment conditions.
What was built
- Lightweight convolutional architecture built in TensorFlow.
- OpenCV preprocessing ahead of training and inference.
- Four-class classification over the retinal fundus dataset.
06Data
Image dataset
- Retinal fundus image dataset across four classes
- Variable image quality and illumination across the set
- Preprocessing applied consistently across train and evaluation splits
07Results
91.14%
Precision
4
Retinal classes
Lightweight
Smaller footprint than standard models
SNPD 2025
Peer-reviewed and published