Development and validation of deep learning algorithms for scoliosis screening using back images

Area of Science:

• Orthopedics
• Medical Imaging
• Artificial Intelligence

Background:

• Adolescent idiopathic scoliosis is a common spinal deformity affecting 0.5-5.2% of adolescents globally.
• Current screening methods often lead to unnecessary referrals and radiation exposure due to low positive predictive values.

Purpose of the Study:

• To develop and validate deep learning algorithms for automated scoliosis screening using unclothed back images.
• To assess the performance of these algorithms compared to human specialists in detecting and grading scoliosis.

Main Methods:

• Development of deep learning algorithms trained on unclothed back images.
• Validation of algorithm accuracy in detecting scoliosis, identifying curves ≥20°, and grading severity.
• Comparative analysis against human specialist performance.

Main Results:

• Deep learning algorithms demonstrated superior accuracy compared to human specialists.
• Algorithms effectively detected scoliosis, identified significant curves (≥20°), and graded severity.
• The approach achieved high accuracy in both binary and four-class classifications.

Conclusions:

• Automated deep learning screening of adolescent idiopathic scoliosis using back images is feasible and accurate.
• This radiation-free method has the potential to improve screening efficiency, reduce costs, and minimize unnecessary referrals.
• The technology can be applied in routine screening and follow-up of patients without exposing them to radiation.