This project was developed as part of the BM2210 – Biomedical Device Design module (3rd semester) with the goal of advancing automated eye disease diagnosis using a smart, AI-powered table-mounted fundoscope. By combining machine learning and IoT, this system enables real-time, accessible, and accurate retinal imaging and classification of diabetic retinopathy severity levels.
We designed a smart table-mounted fundoscope using an ESP32-CAM module paired with a 20D lens to capture high-resolution fundus images. The images are transmitted wirelessly over Wi-Fi for processing using a custom-trained machine learning model.
- Trained on 21,000+ retinal images
- Built using TensorFlow and Keras
- Utilizes a Convolutional Neural Network (CNN) to classify fundus images into 5 levels of diabetic retinopathy:
- Class 0: Normal Fundus
- Class 1: Mild Disease
- Class 2: Moderate Disease
- Class 3: Severe Disease
- Class 4: Proliferative Disease
- Optimized with dropout regularization and careful hyperparameter tuning for strong accuracy.
- Images captured using the ESP32-CAM are transmitted over Wi-Fi for real-time diagnosis.
- Supports remote access, making it ideal for telemedicine applications and use in resource-limited settings.
- Interactive Streamlit dashboard allows clinicians to:
- Upload fundus images
- Get real-time classification results
- View confidence levels for each prediction
- Access reference information for each disease level
This project addresses the increasing need for early detection of diabetic retinopathy, especially in regions with limited access to ophthalmologists. By integrating AI and IoT technologies, our device empowers healthcare providers with faster, data-driven decisions, improving patient outcomes and preventing vision loss.
- Hardware: ESP32-CAM, 20D lens
- Software: Python, TensorFlow, Keras, Streamlit
- Enclosure: SolidWorks
- Communication: Wi-Fi (ESP32 to PC)
- Dataset: A publicly available retinal fundus image dataset (21,000+ samples)
- Add support for mobile diagnostics
- Integrate data logging for patient monitoring over time
- Improve image preprocessing and segmentation for higher model accuracy