Deep learning–based automatic measurement of spinal alignment and implant detection in scoliosis radiographs

Le résumé fourni par la source

Deep learning-based automated analysis offers the potential to streamline workflows, improve reproducibility, and reduce clinician workload. We developed an artificial intelligence (AI) system based on convolutional neural networks to automatically measure spinal alignment and detect spinal implants (pedicle screws and hooks combined as a single category) from scoliosis radiographs. A total of 4585 radiographs from 1671 adolescent idiopathic scoliosis patients across 10 institutions in 2 countries were used to train, validate, and externally test the model. The AI measured coronal and sagittal parameters with mean absolute errors (MAEs) of 2.7° (r = 0.99) for the major curve and 3.7° (r = 0.91) for thoracic kyphosis, regardless of implant presence. Vertebral numbering was performed with an accuracy of 0.97 for recognizing transitional vertebrae. Implant detection achieved high accuracy, with an MAE of 0.18 implants per image (r = 0.99). This comprehensive, multi-institutional validation demonstrates the clinical and research utility of the model in enabling fully automated assessment of spinal deformity. While external validation for coronal preoperative measurements was conducted across four cohorts in four countries, sagittal and postoperative measurements were externally validated at two sites (the United States and Japan); broader multi-region validation of sagittal and postoperative measurements, therefore, remains an important direction for future work.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.