Retinal Anomaly Detection

Explainable A.I model for Retinal Disease Detection using Deep Learning

---

Problem Statement

Early diagnosis of retinal diseases—such as diabetic retinopathy, glaucoma, and macular degeneration—is vital to prevent irreversible vision loss. Traditional manual diagnosis:

• Is time-consuming
• Requires skilled ophthalmologists and if not diagnosed early cand correctly, could lead to irreversible vision loss.
• Can suffer from inter-observer variability

This project builds an automated and explainable AI system to assist healthcare professionals in detecting retinal anomalies accurately and interpretably.

---

My Objective

To build an end-to-end AI pipeline that:

• Detects multiple retinal diseases from fundus images
• Explains its predictions in plain English
• Uses a Grad-Cam (heapmap) to add depth to its prediction

The aim is to make diagnosis faster, more transparent, and educational—for both patients and practitioners.

---

Targeted Diseases

As of now, the diseases covered by the model are:

• Opacity
• Diabetic Retinopathy
• Glaucoma
• Macular Edema
• Macular Degeneration
• Retinal Vascular Occlusion
• Normal Eye

---

Project Overview

Website: https://blueraymusic.github.io/Retinal/

1. Multi-Label Retinal Disease Classifier

• Model: Transferring learning from ResNet50 (changed the 4th layer to train on the data, so as to keep the other trained features)
• Dataset: Fundus images from Kaggle + other public sources
• Labels: Multi-label output for 7 retinal conditions

2. Inference & Upload Interface

• Upload a retinal image via a CLI or web interface
• The model:
  • Verifies image validity
  • Preprocesses and predicts diseases (With Percentages)
  • Displays top diseases with confidence scores
  • Explains the results in plain english covering how and what the eye is subject to

3. Explainability

• Converts technical predictions into human-readable summaries

---

OpenAI Integration

• GPT-4 receives:
  • Predicted diseases and confidence scores
• GPT-4 responds with:
  • A natural language explanation for medical users

---

Tech Stack

Component	Tools / Libraries
Model Training	PyTorch, Torchvision
Data Handling	Numpy
NLP Explanation	OpenAI GPT-4 API / LLama
Inference Server	Streamlit
Optional UI	Html & CSS

---

File Structure

Retinal-Disease-Detection/
│
├── app/
│   ├── pages/
│       └── Retinal_Report.py
│   ├── Analyze_Scan.py         # Main execution script
│   ├── app_main.py             
│   └── app.py                  # Prototype app (v1)
│
├── chart/                      # GLCM chart generation utilities
│   ├── __init__.py
│   ├── glcmchart.py
│   └── utils.py
│
├── data/                       # Datasets & experiment notes
│   ├── test/                   # Test images
│   │   └── ...
│   ├── train/
│   │   ├── train/              # Training images
│   │   │   ├── img.jpg
│   │   │   └── ...
│   │   └── train.csv           # Training metadata
│   ├── analysis.md             # Dataset analysis
│   └── experiment_log.md       # Experiment records
│
├── glcm/                       # GLCM feature extraction & model integration
│   ├── __init__.py
│   ├── glcm.py
│   └── resnet_glcm.py
│
├── ensemble/                   # Ensemble learning scripts
│   ├── ensemble_inference.py
│   └── ensemble_train.py
│
├── visualization/              # Model output & visualization assets
│   ├── output/
│   │   ├── rbgspectrum_plot.png
│   │   └── ...
│   ├── data.png
│   └── ...
│
├── models/                     # Saved models / model notes
│   ├── __init__.py
│   ├── model_checker.py        # checking model existance then download them if not  
│   └── model.md                # Google drive path to the model 
│
├── inference/                  # Inference & explainability tools
│   ├── __init__.py
│   ├── checker.py               # Image validation
│   ├── explainer.py              # Grad-CAM & explanations
│   └── inference.py              # Prediction pipeline
│
├── training_logs/              # Training history & evaluation
│   ├── graphs/                  # Training curve plots
│   │   ├── v1_loss_plot.png
│   │   └── ...
│   ├── prediction_json/         # Saved prediction results
│   │   ├── ensemble_bagging.json
│   │   └── ...
│   ├── v1.txt
│   └── ...
│
├── .gitignore
├── LICENSE
├── index.html                   # Website for the product
├── main.py                      # Main app entry point - Run First
├── train.py                     # Model training script
├── requirements.txt             # Dependencies
├── SECURITY.md                  # Security Readme
├── README.md
└── changelog.md                 # Update history

---

Dataset

• Source:
  Kaggle: VietAI Retinal Disease Detection

• ~3,285 images:

  • 3,210 abnormal
  • 575 normal (Messidor + EyePACS)

• Diseases:

  • Opacity
  • Diabetic Retinopathy
  • Glaucoma
  • Macular Edema
  • Macular Degeneration
  • Retinal Vascular Occlusion (RVO)
  • Normal

Example label structure:

filename	opacity	dr	glaucoma	edema	degeneration	rvo	normal
c24a1b14d253.jpg	0	0	0	0	0	1	0

Data was also accessed via this repo (due to Kaggle auth issues):
zequeira/Retinal-Disease-Detection

---

Sample Output

---

Installation

1. Clone the Repository

git clone https://github.com/username/RetinalAnomalyDetection.git
cd RetinalAnomalyDetection

2. Install Requirements

• Python Version: 3.12.3

pip install -r requirements.txt

3. Application Running

python main.py

---

Current Updates & Future Directions

• Feedback loop for post-diagnosis correction and retraining
• Lightweight offline mode for clinics

-> View Updates Updates.md would contain & track all the experiments advancements and failures regarding the data analysis and model performance. Summarizing it in order to keep track over time.

---

License

MIT License – Free for use in research, education, and development.