TransferLearning

Transfer learning–based chest X-ray classification system using pre-trained CNNs to detect pneumonia vs normal cases, with an end-to-end pipeline including training, FastAPI inference service.

Comment

MedXfer — Transfer Learning for Chest X-Ray Diagnosis

Educational AI project demonstrating how transfer learning can be applied to medical image classification using chest X-ray images. This project covers the full machine learning pipeline: model training, evaluation, backend inference API, and a frontend visualization interface.

Disclaimer This project is for educational and research purposes only. It is not a medical device and must not be used for clinical decision-making.

image

Project Overview

Medical imaging datasets are often small, noisy, and expensive to label. Training deep neural networks from scratch in such settings is inefficient and prone to overfitting.

To address this, transfer learning is used:

  • A deep CNN pre-trained on ImageNet
  • Feature extraction layers reused
  • Final classification layer retrained on Chest X-ray data (NORMAL vs PNEUMONIA)

This project demonstrates:

  • Why transfer learning works
  • How it is implemented in practice
  • How trained models can be served via an API
  • How predictions can be visualized in a frontend UI

Transfer Learning — Conceptual Explanation

Why Transfer Learning?

  • Medical datasets are usually small
  • Training CNNs from scratch requires millions of images

Early CNN layers learn generic visual features such as:

  • edges
  • textures
  • shapes

These features are universal and not domain-specific.

How It Is Applied Here

  • Load a CNN pre-trained on ImageNet (ResNet-18 / ResNet-50)
  • Freeze convolutional layers (feature extractor)
  • Replace the final fully connected layer
  • Train only the new classifier on chest X-ray images
  • Optionally fine-tune deeper layers

This allows the model to:

  • Learn faster
  • Generalize better
  • Avoid overfitting

Project Structure

TransferLearning/
│
├── backend/                 FastAPI inference service
│   ├── app/
│   │   ├── main.py          API entry point
│   │   ├── inference.py    Model loading and prediction
│   │   └── schema.py       Response schemas
│   ├── model/
│   │   ├── model.pt        Trained weights
│   │   └── model_meta.json Model metadata
│   └── requirements.txt
│
├── ml/                      Training and evaluation
│   ├── train.py             Transfer learning training script
│   ├── model_def.py         CNN architectures
│   ├── export_model.py      Export trained model for inference
│   ├── utils.py
│   ├── outputs/
│   │   ├── model.pt
│   │   └── model_meta.json
│   └── requirements.txt
│
├── frontend/                React UI
│   └── src/
│       ├── App.jsx
│       ├── App.css
│       └── api.js
│
├── chest_xray/              Dataset
│   ├── train/
│   ├── val/
│   └── test/
│
└── README.md

Dataset

Chest X-Ray Pneumonia Dataset

Classes:

  • NORMAL
  • PNEUMONIA

Splits:

  • Train
  • Validation
  • Test

Images are resized and normalized using ImageNet statistics to match the pretrained backbone.

Model Architecture

  • Backbone: ResNet-18 (ImageNet pretrained)
  • Input Size: 224 × 224
  • Output: 2 classes (Softmax)
  • Loss: Cross-Entropy
  • Optimizer: Adam
  • Metrics: AUROC, Precision, Recall, F1-Score

Evaluation Results (Test Set)

Metric Value
Accuracy 84.1%
AUROC 0.936
F1-Score (Macro) 0.81

Confusion Matrix:

[[138,  96],
 [  3, 387]]

The model shows:

  • High sensitivity for pneumonia detection
  • Acceptable trade-off between precision and recall

Backend (FastAPI)

The backend provides:

  • Model loading
  • Image preprocessing
  • Prediction endpoint

Available Endpoints

Method Endpoint Description
GET /health API health check
GET /model-info Model metadata
POST /predict Predict X-ray image

Predictions return:

  • Predicted label
  • Confidence score
  • Full class probability distribution

Frontend

The frontend is a React-based interface that:

  • Allows image upload
  • Sends images to backend API
  • Displays prediction results visually
  • Explains the transfer learning concept in text form

UI focus:

  • Large layout (no small cards)
  • Clear medical-style design
  • Left-aligned explanation panel
  • Right-aligned prediction panel

Running Locally

Train Model

cd ml
pip install -r requirements.txt
python train.py --data_dir ../chest_xray

Export Model

python export_model.py

Run Backend

cd backend
pip install -r requirements.txt
uvicorn app.main:app --reload

Run Frontend

cd frontend
npm install
npm run dev

Learning Outcomes

This project demonstrates:

  • Practical transfer learning implementation
  • Medical image preprocessing
  • Model evaluation using clinical metrics
  • AI system integration (ML → API → UI)
  • End-to-end ML system design

Author

Hadi Mostafa Computer Engineering Student

Focus areas:

  • Artificial Intelligence
  • Computer Vision
  • Medical AI Systems

License

This project is released for educational and academic use only.

Built With

Share this project:

Updates