PRELIMINARY RESULTS OF CLINICAL TESTING OF A NEW INTELLECTUAL PROGRAM ESPER.SCOLIOSIS FOR AUTOMATIC DIAGNOSIS OF SCOLIOSIS ON FRONTAL RADIOGRAPHS

Abstract

Introduction. Scoliosis is a lateral curvature of the spine with torsion of the vertebral bodies and their posterior elements. Cobb’s angle is considered the “gold standard” for diagnosis of scoliosis on frontal radiographs and it is defined by the most tilted vertebrae toward the apex of the curve (end vertebrae). To measure Cobb’s angle the doctor, have to apply manually lines parallel to the end vertebrae. Variability of Cobb’s angle measurements is a current disadvantage of this method. Esper.Scoliosis is an automatic program that is based on machine learning. The aim of the study is to confirm that using program Esper. Scoliosis is in keeping with accuracy metrics. It can detect vertebrae form TI–LV, as well as four points of each vertebral body, according to which the program can construct the angle and determine the degree of scoliosis automatically. Materials and methods. 120 X-ray images were collected retrospectively from the archive of St. Petersburg State Medical Institution “Mariinskaya hospital” and the Federal Scientifi c Centre of Rehabilitation of the Disabled n.a. G.A. Albrecht. The data set consists of 60 images without scoliosis and 60 images with different grades of scoliosis from 1 to 4 that were distributed evenly (15 X-ray for each grade). Results. The results of the index test (the new program) were compared to a reference test (radiologist conclusion), and a standard set of metrics was calculated (sensitivity, specificity, accuracy and the area under ROC curve). Conclusion. Our study shows that Esper.Scoliosis system has a high accuracy in diagnosing scoliosis (ROC AUC = 0.9). Based on this clinical test, it is recommended to use program Esper.Scoliosis in clinical practice as an objective tool for determining the degree of scoliosis.