Radiographic Analysis of Scoliosis Using Convolutional Neural Network in Clinical Practice.

Abstract

Purpose: To assess the reliability and accuracy of an automated Cobb angle measurement (ACAM) using a convolutional neural network (CNN) for scoliosis evaluation and to compare measurement times.

Materials and methods: ACAM was applied to spine radiographs in 411 patients suspected of scoliosis. Observer 1 (consensus of two musculoskeletal radiologists) and observer 2 (a radiology resident) measured Cobb angle (CA). CA measurements were categorized using observer 1's measurements as the reference standard. Inter-observer reliability and correlation were assessed using intraclass correlation coefficient (ICC) and Spearman's rank correlation coefficient, respectively. Accuracy and measurement time of ACAM and observers were evaluated.

Results: ACAM demonstrated excellent reliability and very high correlation with observer 1 (ICC = 0.976, Spearman's rank correlation = 0.948), with a mean CA difference of 1.1. Overall accuracy was high (88.2%), particularly in mild (92.2%) and moderate (96%) scoliosis. Accuracy was lower in spinal asymmetry (77.1%) and higher in severe scoliosis (95%), although the CA was lower compared to the observers. ACAM significantly reduced measurement time by nearly half compared to the observers (p < 0.001).

Conclusion: ACAM using CNN enhances CA measurement for assessing mild or moderate scoliosis, despite limitations in spinal asymmetry or severe scoliosis. Nonetheless, it substantially decreases measurement time.