Cervical Spine Fracture Classification using Deep Learning

This project presents a computer-aided diagnosis system for classifying cervical spine fractures using deep learning models. The system leverages AlexNet and GoogleNet architectures to identify and classify cervical spine injuries as normal, fracture, or dislocation. With an accuracy of 99.67%, the proposed model outperforms the average radiologist's accuracy, providing a reliable solution for assisting in the diagnosis of critical cervical spine injuries.

Overview

Cervical spine fractures are often associated with severe consequences, such as paralysis or death. These injuries require timely and accurate diagnosis to prevent further complications. Our deep learning model is designed to assist doctors by automating the classification of cervical spine X-ray images, reducing human error and enabling quicker decision-making.

The dataset used contains 772 cervical spine fractures and 707 normal images. Our model, trained using AlexNet and GoogleNet with transfer learning, classifies X-ray images into three categories: normal, fracture, and dislocation. The final accuracy achieved is 99.67%, higher than the radiologists’ average accuracy of 90-95%.

Key Features

• Deep Learning Models: AlexNet and GoogleNet are used for feature extraction and classification.
• Saliency Maps: Visualization of model attention to ensure explainability in the classification process.
• High Accuracy: Achieves an accuracy of 99.67%, surpassing traditional radiological methods.
• Transfer Learning: Fine-tuning of pre-trained AlexNet and GoogleNet models for cervical spine classification.
• Efficient Performance: Designed to run on standard PCs or embedded systems with minimal setup.

Methodology

1. Dataset:

   • Images are divided into three categories: normal, fracture, and dislocation.
   • Preprocessing includes resizing images and converting them into a format suitable for AlexNet and GoogleNet input layers.

2. Model Architecture:

   • AlexNet: Composed of five convolutional layers followed by three fully connected layers, using ReLU activations and SoftMax for final classification.
   • GoogleNet (Inception): Utilizes different-sized convolutions (5x5, 3x3, 1x1) to capture multi-scale features.

3. Training:

   • Dataset is split into 70% training, 15% validation, and 15% testing.
   • Optimization is done using Stochastic Gradient Descent (SGD) with a learning rate of 0.001 and scheduled annealing to prevent local minima.
   • Performance metrics include accuracy, precision, recall, F1-score, and ROC curves.

Results

• Model Performance: Achieved 99.67% accuracy, significantly higher than radiologists (92%).
• Training: The model was trained over 30 iterations with each batch containing 10 samples.
• Saliency Maps: Used to confirm the model's focus on key features of the X-rays that correspond to fractures or dislocations.

Dataset

The dataset used for this project is available on Kaggle and contains:

• 530 Cervical Spine Dislocation Images
• 772 Cervical Spine Fracture Images
• 707 Normal Images

Performance Metrics

Metric	Value
Accuracy	99.67%
Sensitivity	76%
Specificity	97%
Precision	99.6%
F1-Score	99.6%

License

This project is licensed under the MIT License - see the LICENSE file for details.