Classification of spinal curvature types using radiography images: deep learning versus classical methods

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Spinal Deformity Analysis

Background:

• Scoliosis presents with C-shaped or S-shaped spinal curves, challenging to diagnose due to vertebral equilibrium timing, observer bias, and image quality.
• Accurate classification of spinal curvature type is crucial for effective scoliosis management.

Purpose of the Study:

• To evaluate spinal deformity by automatically classifying scoliosis curvature types.
• To compare the performance of classical machine learning algorithms with deep learning models for spinal curvature classification.

Main Methods:

• Utilized Support Vector Machine (SVM) and K-Nearest Neighbors (KNN) algorithms with low-level (texture) and local patch-based (Bag of Words) features.
• Employed pre-trained deep networks (Xception, MobileNetV2) with SVM as the final activation function for automated feature extraction.
• Trained models on a private dataset of 1000 anterior-posterior (AP) spine radiographic images using transfer learning.

Main Results:

• Pre-trained deep networks demonstrated approximately 10% higher accuracy in classifying C-shaped versus S-shaped spinal curvatures compared to classical methods.
• Automated feature extraction via deep networks, particularly Xception and MobileNetV2 with SVM, outperformed traditional feature extraction techniques combined with SVM/KNN.

Conclusions:

• Pre-trained deep learning networks with SVM integration offer a more accurate and automated approach for classifying scoliosis spinal curvature types.
• This AI-driven method enhances diagnostic capabilities for spinal deformities, potentially overcoming limitations of manual assessment and traditional machine learning.