Pneumonia Detection with ResNet50 and FastAPI

This project implements a deep learning model for detecting pneumonia from chest X-ray images using a fine-tuned ResNet50 architecture with focal loss.
It includes both the training pipeline and a REST API built with FastAPI for real-time predictions.

---

Project Overview

• Data preprocessing: Image augmentation and normalization with ImageDataGenerator
• Model architecture: Pretrained ResNet50 (imagenet weights) with custom classification head
• Loss function: Custom focal loss to handle class imbalance
• Training: Transfer learning with partial freezing of base model layers
• Evaluation: Precision, recall, F1-score, confusion matrix, and threshold tuning
• Deployment: FastAPI server exposing a /predict endpoint for image input

---

Requirements

• Python 3.8+
• TensorFlow
• FastAPI
• Uvicorn
• NumPy, Pillow
• scikit-learn, matplotlib, seaborn

Install dependencies with:

pip install tensorflow fastapi uvicorn numpy pillow scikit-learn matplotlib seaborn

---

Dataset

Place your dataset folder with the following structure inside the specified path (e.g., C:\Users\Yasaman\Desktop):

chest-xray-pneumonia-project/
├── train/
│   ├── NORMAL/
│   └── PNEUMONIA/
├── test/
│   ├── NORMAL/
│   └── PNEUMONIA/

---

Training the Model

Run the training script (e.g., train.py) that includes:

• Data loading and augmentation
• Model definition with ResNet50 and focal loss
• Callbacks for early stopping and learning rate reduction
• Saving the best trained model as myy_model.keras

python train.py

---

Running the API Server

Start the FastAPI server with:

uvicorn main:app --reload

The API will be available at: http://127.0.0.1:8000

---

Using the API

Send a POST request with a chest X-ray image file to /predict endpoint. Example using curl:

curl -X POST "http://127.0.0.1:8000/predict" -F "file=@path_to_image.jpg"

Response example:

{
  "prediction": "Pneumonia",
  "confidence": 0.912
}

---

Notes

• The model uses a threshold tuned based on F1-score for better classification accuracy.
• The API expects RGB images resized to 224x224 pixels.

---

License

This project is released under the MIT License.

If you want, I can help you generate a minimal training script file and the API main file separately too!