Predicting Nonidiopathic Scoliosis from Plain Radiographs: A Deep-Learning Approach.

Background: Scoliosis has various etiologies, ranging from idiopathic, congenital, to those with spinal cord abnormalities. Advanced imaging is not always feasible and limited by cost, low-yield, and anesthesia risks in children. Given artificial intelligence's potential, we hypothesize that a deep learning (DL)-based image classifier will demonstrate superior performance in radiographic classification of scoliosis etiology compared witht experienced spine surgeons based on spine radiographs. Methods: One thousand thirty-six pediatric patients from single institution with scoliosis diagnosis and paired anterior-posterior - lateral images were included. Patients were manually classified based on their scoliosis etiology from chart review, including previous spine imaging studies when available. Categories included idiopathic, congenital, and spinal cord pathology. Images were randomized and assigned for training, validation, and testing. A DL-classifier using EfficientNet B4 architecture was trained on the radiographs. Accuracy and positive predictive value as defined by precision, recall, and F1-score were calculated to assess final performance metrics. Results: The trained classifier performed well at identifying correct etiologies, with F1-Score (harmonic mean of precision-recall) of 0.97. Model precision was 0.99 for adolescent idiopathic scoliosis, 0.89 for congenital, and 0.78 for spinal cord pathology. Performance was higher on more common classes, with lower performance observed in class with fewer images. The algorithm has highest overall precision (0.96), recall (0.96), and F1 (0.96), while Surgeons 1 and 2 have lower accuracies (precision 0.80, recall 0.79, and F1 0.79 and precision 0.76, recall 0.67, and F10.71, respectively). There was no clear pattern identified regarding the surgeons' errors and poor agreement on which X-rays corresponded to underlying spinal cord pathology. Conclusion: A DL-convolutional-neural-network classifier has been trained to high degree of accuracy to distinguish between 3 scoliosis etiologies on pediatric spine radiographs. This model could provide a novel tool to clinicians to decide when to refer for axial imaging. It had superior diagnostic performance compared with experienced spine surgeons. After further validation and refinement, a clinician could put a patient's X-rays into the predictive tool for analysis and use the results to discuss with the family the risk of finding spinal cord pathology, helping them determine together the best timing for an MRI. Level of Evidence: Level III. See Instructions for Authors for a complete description of levels of evidence.