JobayerFaisal/Retinal_Disease_Detection

🩺 OCT Retinal Disease Classification using ResNet50 + LSTM

📌 Overview

This project implements a deep learning model for Optical Coherence Tomography (OCT) retinal disease classification using a hybrid architecture:

• —ResNet50 (feature extraction)
• —LSTM (spatial sequence modeling)
• —Fully Connected layer (multi-class classification)

The model classifies OCT images into the following categories:

• —AMD
• —CNV
• —CSR
• —DME
• —DR
• —DRUSEN
• —MH
• —NORMAL

The application is deployed using Streamlit and hosted on Hugging Face Spaces.

🧠 Model Architecture

Backbone

• —Pretrained ResNet50 (final classification layers removed)
• —Extracts 2048-dimensional feature maps

Sequence Model

• —LSTM layer

• —Input size: 2048
• —Hidden size: 256
• —Number of layers: 1

Output Layer

• —Fully connected layer
• —Output classes: 8

Architecture Flow:

Input Image (224x224)
        ↓
ResNet50 (Feature Extractor)
        ↓
Reshape to sequence
        ↓
LSTM
        ↓
Fully Connected Layer
        ↓
Softmax Prediction

📂 Project Structure

OCT-Classification-App/
│
├── app.py              # Streamlit application
├── model.py            # Model architecture definition
├── model.pth           # Trained model weights
├── requirements.txt    # Dependencies
└── README.md           # Project documentation

⚙️ Installation (Local Setup)

1️⃣ Clone Repository

git clone https://huggingface.co/spaces/YOUR_USERNAME/YOUR_SPACE_NAME
cd YOUR_SPACE_NAME

2️⃣ Create Virtual Environment (Recommended)

python -m venv venv
source venv/bin/activate   # Mac/Linux
venv\Scripts\activate      # Windows

3️⃣ Install Dependencies

pip install -r requirements.txt

▶️ Run Locally

streamlit run app.py

App will open at:

http://localhost:8501

📦 Deployment (Hugging Face Spaces)

1. 1.Create a new Space
2. 2.Choose:

• —SDK: Streamlit
• —Hardware: CPU Basic
• —Upload:

• —app.py
• —model.py
• —model.pth
• —requirements.txt
• —Commit & Deploy

No additional configuration required.

🔄 Preprocessing Pipeline

During inference, each image undergoes:

• —Resize to 224×224
• —Convert to Tensor
• —ImageNet normalization

transforms.Normalize(
    mean=[0.485, 0.456, 0.406],
    std=[0.229, 0.224, 0.225]
)

This ensures consistency with ResNet50 expectations.

🖥️ Requirements

streamlit
torch
torchvision
pillow
numpy

📊 Model Training Details

• —Loss Function: CrossEntropyLoss
• —Optimizer: Adam
• —Input size: 224×224 RGB images
• —Training Framework: PyTorch

Model weights are saved using:

torch.save(model.state_dict(), "model.pth")

⚠️ Important Notes

• —The model must be loaded with weights=None for ResNet50 during deployment.
• —Ensure architecture in model.py exactly matches training architecture.
• —CPU inference supported.

📈 Future Improvements

• —Add confidence scores
• —Add Grad-CAM visualization
• —Add model performance metrics display
• —Convert to TorchScript for faster inference
• —Add API endpoint (FastAPI version)

👤 Author

Jobayer Faisal Fahim Machine Learning & AI Researcher