A Validated Capsule Network to Predict Curve Progression in Adolescent Idiopathic Scoliosis Based on Posteroanterior X-rays at First Visit

Abstract

Introduction: Early curve progression risk prediction is essential for the management of adolescent idiopathic scoliosis (AIS). Prior studies have revealed potential predictive value of 3D morphological parameters for curve progression, but acquisition of parameters rely on specialised biplanar imaging equipment and time-consuming reconstruction. This study aimed to formulate a deep learning model with standing posteroanterior (PA) X-rays as input in the distinguishment of progressive group (P) and non-progressive group (NP) at first clinical visit. Methods: This is a retrospective study consisting of a training cross-validation cohort (328 AIS patients), an independent testing cohort (110 AIS patients) utilising EOS images and a cross-platform validation cohort (52 AIS patients) upon standard standing PA X-ray projections. Results: The predictive model achieved an accuracy of 76.4%, a sensitivity of 74.5% and a specificity of 80% on independent testing cohort (110 cases, 55 P and 55 NP). The model cross-platform validation (52 cases, 24 P and 28 NP) upon standard standing PA X-ray projections achieved an accuracy of 76.9%, a sensitivity of 70.8% and a specificity of 82.1%. Discussion and Conclusion: This is the first attempt at automated prediction of AIS curve progression based on radiomics and deep learning, towards directing treatment strategy at first visit. The model takes PA X-rays as input at AIS patients first visit and classifies patients as P or NP subjects. It could help to recommend timely clinical decisions on bracing treatment for the potential progressive group and to avoid over treatment of the likely non-progressive patients.