TensorFlow-based Eye Disease Detection AI System: From Data Preprocessing to Production-Level Deployment

This project is a professional eye disease detection AI system based on TensorFlow and Keras, using the EfficientNetB0 architecture and transfer learning technology to achieve four-class image classification of normal fundus, glaucoma, diabetic retinopathy, and cataracts. The project includes a complete training pipeline, evaluation tools, and a CLI interface. It is clearly labeled as an engineering research project, and prediction results require review by clinical doctors, reflecting the safety principles of medical AI.

• Original author/maintainer: erencankur
• Source platform: GitHub
• Original title: Eye_Disease_Detection
• Original link: https://github.com/erencankur/Eye_Disease_Detection
• Release date: 2026-06-12

The project is positioned as a medical imaging AI application case, focusing on automatic detection of eye diseases. It is not a diagnostic medical device, and all results require review by doctors.

Model Selection: Uses the pre-trained EfficientNetB0 model, balancing accuracy and efficiency through compound scaling; transfer learning is divided into two stages: freezing the main part for warm-up training, and fine-tuning the entire network with a low learning rate.

Data Engineering: Uses stratified sampling to split the dataset (maintaining class proportions); builds an efficient data pipeline using the tf.data API (batching, prefetching, parallel loading); the dataset comes from public fundus images on Kaggle, and a Kaggle download script is provided to simplify acquisition.

Training Strategy: Handles imbalance with class weights; uses early stopping to monitor validation set performance to prevent overfitting; saves the best model with checkpoints; records training metrics with TensorBoard.

Project Structure: Core modules include cli.py (command-line interface), config.py (configuration management), data.py (data processing), model.py (model construction), etc.; complete documentation (MODEL_CARD.md, SYSTEM_DESIGN.md); configures unit tests and code quality tools (ruff, pytest).

Local Training: Isolates development with a virtual environment; starts training via CLI: eye-disease train --config configs/default.yaml, output includes model files, class mappings, evaluation metrics, etc.

Inference Prediction: Single image prediction command: eye-disease predict --image path/to/image.jpg --model ... --top-k 3, supports top-k result output.

Safety Design: Confidence threshold mechanism (low confidence reminder); all outputs include a disclaimer, emphasizing the need for clinical review.

• Modularity: Breaks down functions into independent modules; config-driven design facilitates hyperparameter experiments;
• Engineering: Mature practices such as virtual environments, typed configurations, and unit tests;
• Medical AI Considerations: Model card writing, safety note annotations, confidence reporting mechanism, reflecting the principle of human-machine collaboration.

Limitations: Limited dataset size; only supports four-class binary classification (normal/abnormal), not covering multi-label classification needs.

Improvement Directions: Introduce more data sources for pre-training; try new architectures like Vision Transformer; support multi-label classification; develop model interpretation tools like Grad-CAM.

Eye_Disease_Detection is a well-structured and fully documented eye AI project that demonstrates the full process of building an image classification system with TensorFlow, embodies engineering thinking and safety awareness for medical AI, and provides a good reference case for medical imaging AI developers.