Fracture Detection in X-ray Images

A machine learning project to classify fractures in X-ray images using transfer learning and advanced fine-tuning techniques.

---

📖 Table of Contents

• 📁 Project Overview
• 📊 Dataset
• ⚙️ Workflow
• 💻 Tools and Technologies
• 🧠 Model Architecture
• 📈 Results
• 🚀 Future Improvements

---

📁 Project Overview

This project aims to detect fractures in X-ray images by leveraging a pre-trained ResNet50 model through transfer learning and fine-tuning. The goal is to assist medical professionals in identifying fractures efficiently and accurately.

• Problem: Detect fractured vs. non-fractured bones.
• Solution: Develop a binary classification model using advanced machine learning techniques.
• Objective: Achieve high accuracy and robust generalization on unseen X-ray images.

---

📊 Dataset

• Source: Custom X-ray dataset with two categories: fractured and non-fractured.
• Size:
  • Training: ~4000 images.
  • Validation: ~1000 images.
  • Test: ~1000 images.
• Preprocessing:
  • Grayscale conversion to 3 channels for compatibility with pre-trained models.
  • Normalization to match ImageNet statistics.
  • Augmentations: Random cropping, flipping, resizing.

---

⚙️ Workflow

1. Data Preparation:

   • Normalized, resized, and augmented the dataset.
   • Ensured balanced class distribution.

2. Exploratory Data Analysis (EDA):

   • Visualized dataset distribution and augmented samples.
   • Checked for class imbalance.

3. Model Training:

   • Used ResNet50 pre-trained on ImageNet.
   • Fine-tuned specific layers while freezing others.
   • Applied Early Stopping and learning rate scheduling.

4. Evaluation:

   • Tested on unseen data.
   • Computed metrics: Accuracy, F1-score, Precision, Recall.
   • Visualized confusion matrices and performance curves.

---

💻 Tools and Technologies

• Programming Language: Python 🐍

• Libraries:

  • PyTorch (Model building and training)
  • torchvision (Transfer learning, preprocessing)
  • scikit-learn (Metrics and evaluation)
  • matplotlib, seaborn (Visualizations)

• Hardware:

  • GPU-enabled system (NVIDIA CUDA) for efficient training.

---

🧠 Model Architecture

• Base Model: ResNet50
  • Pre-trained on ImageNet.
  • Final fully connected layer modified to output 2 classes.
• Optimizer: Adam
• Loss Function: CrossEntropyLoss
• Learning Rate Scheduler: StepLR (decays every 7 epochs).

---

📈 Results

Metric	Original Model (Frozen)	Fine-Tuned Model (Unfrozen)
Accuracy	0.80	0.99
F1-Score	0.80	0.99
Precision	0.80	0.985
Recall	0.80	0.985

Confusion Matrix (Fine-Tuned Model)

---

🚀 Future Improvements

1. Augmented Dataset: Include more samples to improve generalization.

---

📂 Project Structure

📦 Fracture-Detection-Project  
├── 📁 data/                # Contains train/val/test datasets  
├── 📁 notebooks/           # Jupyter notebooks for training and EDA  
├── 📁 models/              # Saved model weights (e.g., model.pth)  
├── 📁 results/             # Evaluation metrics and visualizations  
└── README.md               # Project README