Radiographic Analysis of Scoliosis Using Convolutional Neural Network in Clinical Practice

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Orthopedics

Background:

• Scoliosis evaluation relies on accurate Cobb angle (CA) measurements from spinal radiographs.
• Manual CA measurement can be time-consuming and subject to inter-observer variability.
• Automated methods are being developed to improve efficiency and consistency.

Purpose of the Study:

• To evaluate the reliability and accuracy of an automated Cobb angle measurement (ACAM) system utilizing a convolutional neural network (CNN).
• To compare the measurement times of ACAM against human observers for scoliosis assessment.
• To establish ACAM's performance across different scoliosis severity levels and spinal conditions.

Main Methods:

• ACAM was applied to 411 spinal radiographs of patients with suspected scoliosis.
• Cobb angle measurements were performed by ACAM and two human observers (radiologists).
• Inter-observer reliability, correlation, accuracy, and measurement times were statistically analyzed using ICC and Spearman's rank correlation.

Main Results:

• ACAM demonstrated excellent reliability (ICC=0.976) and high correlation (Spearman's=0.948) with expert measurements, with a mean difference of 1.1 degrees.
• High overall accuracy (88.2%) was observed, with superior performance in mild (92.2%) and moderate (96%) scoliosis.
• ACAM significantly reduced measurement time by approximately 50% compared to human observers (p < 0.001).

Conclusions:

• Automated Cobb angle measurement using CNNs provides a reliable and accurate tool for assessing mild to moderate scoliosis.
• While accuracy is lower for spinal asymmetry and severe scoliosis, ACAM substantially decreases evaluation time.
• ACAM shows promise in enhancing the efficiency of scoliosis diagnosis and monitoring.