Enhanced Vision Transformer with Custom Attention Mechanism for Automated Idiopathic Scoliosis Classification

Area of Science:

• Orthopedics and Rehabilitation Medicine
• Artificial Intelligence in Healthcare
• Medical Imaging Analysis

Background:

• Scoliosis is a complex spinal deformity impacting posture and health, often graded using the subjective Cobb angle measurement from X-rays.
• Manual Cobb angle calculation is time-consuming and prone to inter-observer variability, necessitating objective and efficient diagnostic tools.
• Deep learning models offer potential for automated and objective assessment of spinal deformities.

Purpose of the Study:

• To develop and evaluate an enhanced Vision Transformer (ViT) model for objective and rapid scoliosis diagnosis.
• To improve the accuracy and efficiency of Cobb angle assessment in scoliosis classification.
• To provide clinicians with a reliable AI-driven tool for evaluating scoliosis severity.

Main Methods:

• An enhanced ViT architecture incorporating a custom attention mechanism was proposed for scoliosis classification.
• A dataset comprising 7 classes from 1456 patients was utilized for model training and validation.
• The performance of the proposed model was benchmarked against ResNet50, Swin Transformer, and standard ViT models.

Main Results:

• The enhanced ViT model achieved a classification accuracy of 95.21% in diagnosing scoliosis.
• The proposed architecture demonstrated superior performance compared to ResNet50, Swin Transformer, and standard ViT.
• The custom attention mechanism in the ViT model contributed to improved diagnostic accuracy.

Conclusions:

• The enhanced ViT model offers a highly accurate and objective method for scoliosis diagnosis.
• This AI-driven approach can significantly reduce diagnostic time and subjectivity in clinical practice.
• The developed model shows promise for widespread adoption in orthopedic and rehabilitation settings for spinal deformity assessment.