A fully autonomous AI system for accurate and reproducible Cobb angle measurement in adolescent idiopathic scoliosis: a multicenter study

BACKGROUND CONTEXT: The gold standard diagnostic test for Adolescent idiopathic scoliosis (AIS) involves manually measuring the spinal column deformity by determining the Cobb angle on a full-spine X-ray image. This measurement involves a subjective interpretation of vertebrae position and angle calculation, and inter-observer variability is widely accepted as one of the main causes of diagnosis uncertainty. PURPOSE: Our objective was to develop an automated and reproducible system based on artificial intelligence (AI) to assist in Cobb angle estimation on full-spine radiographs without human intervention. STUDY DESIGN: Retrospective, observational, multicenter study. METHODS: We performed a multicenter study involving 4 tertiary hospitals in which we collected full-spine anteroposterior/posteroanterior (AP/PA) X-ray images from AIS patients with Cobb angles ranging from mild to severe. Images were analyzed by 3 independent clinicians in each center (first dataset). Any discrepancies in clinician-reported measurements prompted reevaluation of images and data curation. We developed a deep learning pipeline featuring two specialized AI models designed to detect the spine's curvature from X-ray images, identify the individual vertebrae, and accurately estimate the Cobb angles of all curves detected in the spine. RESULTS: From a total of 484 X-ray images collected, spine surgeons reached consensus on 1,054 curves. Initial analysis identified 86.4% of these curves, with a mean absolute error (MAE) of 2.41°±3.24° relative to the consensus measurement after reevaluation and with error values ranging from -1.30° to 40.7°. In comparison, our SPinal Autonomous Radiological Cobb-assessment (SPARC) AI system detected 94.0% of the consensus curves, with a MAE of 3.01°±2.71°, which is within the clinical acceptance threshold (≤6°), and with a more constrained range of error showing values from -14.6° to 20.3°. CONCLUSION: SPARC is an AI-based system developed for automatic, reproducible, and accurate calculation of Cobb angles in full AP/PA spine radiographs without human intervention. SPARC demonstrates superior performance by detecting a higher proportion of spinal curves (94.0% vs 86.4%) and achieving a lower error range in Cobb angle estimation (±20.3° vs ±41.3°) compared to the initial evaluation by 3 specialists with more than 10 years' experience. CLINICAL SIGNIFICANCE: SPARC removes the intraobserver error and inter-observer variability inherent to manual measurements, and significantly decreases radiograph measurement and interpretation times, thus supporting clinicians in patient management and providing a reliable tool for less experienced practitioners involved in the care of patients at all stages of the AIS journey.