Validation of a decision support system for determining the severity of scoliotic spinal deformity using radiographic image analysis

Objective. To prove the possibility of using a domestic computer program in clinical practice to determine Cobb angle by means of comparative analysis of the obtained automated data with the data of manual measurement by specialists. Material and Methods. A total of 411 digital x-rays of the spine of children and adolescents were selected from the medical database of Prosthetic and Orthopedic Center «Scоliologic.ru». They were measured by a radiologist with significant experience in vertebrology (VR-standard), by a radiologist without experience in vertebrology (R-beginner) and a computer program (CP). The CP data were compared with the standard twice – initially (CP1) and after fine-tuning (CP2). The mean absolute error and mean absolute deviation of the standard data of Cobb angle measurements were analyzed when compared with the indicators obtained by R-beginner, CP1 and CP2 for different types of scoliosis according to the Rigo classification, and in determining the main curve of different magnitude from 20° to 41° and more. The Pearson coefficient (R) and the intraclass correlation coefficient (ICC) were calculated. Results. After fine-tuning, the domestic computer program improved the accuracy of measurement in general for curves and types of scoliosis, exceeding the R-beginner indicators almost twice in mean absolute error. The previously identified program drawback in measuring the magnitude of the lumbar (lumbosacral) curve was eliminated. The CP2 data have the highest correlation with the standard (R = 0.94). The excellent level of reliability of the program (ICC = 0.95 when counting on the main curve and 0.97 when counting on all curves) comparable with foreign analogues was proved. It was also confirmed that the average absolute deviation of ±3.2° and ±4.0° for the main curve corresponds to foreign data. Conclusion. It is possible to conclude that the domestic computer program may be validated, since it has been proven that when compared with a reference measurement, its current algorithm provides accuracy higher than that of a radiologist with no experience in vertebrology, and is comparable with foreign analogues.