Retinal Disease Detection for the Second largest Ophthalmology Clinic in Asia, backed by Crys Capital
Business Problem
About Retinal Image Extraction
Retinal diseases such as Cataract, Diabetic Retinopathy, and Glaucoma are leading causes of vision impairment worldwide. Early detection and classification of these diseases play a crucial role in timely medical intervention. This project leverages Convolutional Neural Networks (CNNs) to build an automated image classification system capable of identifying and categorizing retinal diseases into four classes: Cataract, Diabetic Retinopathy, Glaucoma, and Normal Retina. The solution involves training and deploying two CNN architectures, SmallRetinaNet and RetinaNet, to achieve efficient and accurate disease detection.
Solution
To address these challenges, an advanced solution leveraging cutting-edge technologies was developed. The solution focuses on precision, scalability, and usability to empower healthcare providers and improve clinical outcomes.
1.
SmallRetinaNet (Lightweight Model): Three convolutional layers (16, 32, 64 filters) with Batch Normalization and LeakyReLU activations.
4.
High feature extraction: Max pooling for down sampling, fully connected layers with dropout to prevent overfitting.
7.
Programming Language: Python.
10.
Data Augmentation: Image flipping, rotation, contrast adjustment.
2.
Fast inference: Max pooling for spatial dimension reduction, fully connected layers with dropout for regularization. Optimized for mobile or low-resource deployment.
5.
Deployment & Scalability: TensorFlow Serving, Flask API for integration into hospital systems.
8.
Data Preprocessing: OpenCV, NumPy, Pandas.
3.
RetinaNet (High-Capacity Model): Four convolutional layers (16-128 filters) with Batch Normalization and ReLU activations.
6.
Deep Learning Framework: TensorFlow / PyTorch.
9.
Model Training & Optimization: Adam optimizer, Cross-entropy loss function.
Architecture
