JMIR Publications

ABSTRACT

Background:

Scoliosis is a spinal deformity in which one or more segments of the spine bend to the side or are accompanied by vertebral rotation.

Objective:

We attempt to evaluate the reliability and accuracy of the new artificial intelligence application based on machine learning to automatically measure the Cobb angle of scoliosis.

Methods:

We used the bland-Altman test analyzed the differences between the results of Cobb angle measured by the application and the picture archiving and communication systems (PACS). We evaluated the relationship between the results from the application and the PACS by spearman and linear regression.

Results:

Among 601 children with scoliosis, 89 were male and 512 were female, aged 10-17 years. The average absolute error of Cobb angle measured by two functions of the application was 2.00 and 2.08 The 95% LoA of Cobb angle measured by two functions of the application was - 4.7~4.9 and - 4.9~4.9 respectively. The 95% LoA of Cobb angle in patients with mild scoliosis was - 4.3~4.6 and - 4.4~4.7 respectively, and the measurement accuracy was 97.9%. The 95% limit of agreement (LoA) of moderate scoliosis was - 4.9~5.2 and - 5.1~5.1, respectively, and the measurement accuracy was 97.1%. The 95% LoA of severe scoliosis was - 5.2~5.0 and - 6.0~4.8 respectively, and the measurement accuracy was 92.9%. The cobb angle measured by three observers twice before and after using photo shooting function has good repeatability (ICC was 0.996, 0.996, 0.997 respectively). The consistency between the observers was excellent (ICC was 0.997).

Conclusions:

The new artificial intelligence platform is accurate and repeatable in the automatic measurement of the Cobb angle of the main curvature of scoliosis.