Prédiction de l’évolution de la scoliose idiopathique de l’adolescent à partir d’images radiographiques biplanaires

Abstract

Adolescent Idiopathic Scoliosis (AIS) affects 1 to 4 % of adolescents. Regular radiographic follow-up of these adolescents allows to observe the progression of the pathology and to propose an adapted therapeutic management at the appropriate time. In recent years, several authors have proposed methods to predict the evolution of AIS from the first diagnostic radiograph, in order to reduce exposure to X-rays and to anticipate the therapeutic approach. This is the case of our clinical partner, CHU Sainte Justine, which published a method for predicting the evolution of AIS by linear regression from clinical parameters obtained from the diagnostic bi-planar radiographs (Nault et al., 2020). Most methods for predicting the evolution of AIS do not consider the influence of brace wearing during adolescent growth. Thus, this work proposes first to evaluate the prediction method of our clinical partner, and then to develop a method for predicting the evolution of AIS considering the wearing of a brace during growth. The method of Nault et al., (2020) allows the prediction of the amplitude of the spinal deformation (Cobb angle) at the bone maturity of the adolescent. Its evaluation consisted of estimating the inter-operator reproducibility of its clinical parameters and their impact on prediction, validating the data used for its design, and testing the method on a new cohort. The four clinical parameters from the three-dimensional reconstruction of the diagnostic radiograph used in the method, the Cobb angle, the plane of maximum curvature angle, and the intervertebral disc wedging of T3-T4 and T8-T9 have an inter-operator 2RMSsd reproducibility of 4.7°, 4.9°, 3.7°, 4.2°, generating a variability of the predicted Cobb angle of 2.43°, 3.44°, 3.25° and 0.61°, respectively. The three-dimensional spine reconstructions used in the design of the method were verified. A high variability of the clinical parameters between the initial reconstruction and the verification was observed (4.5°, 3.3°, 5.5° and 16.8° for these same parameters). The method suggested a coefficient of determination R² of 0.643 in training (n = 172). Its test on a new cohort (n = 127) demonstrates an R² coefficient of determination of 0.329 and a mean absolute difference of 7.8° between the predicted and actual amplitude of curvature at bone maturity. The evaluation of our clinical partner's method suggests that it has limited generalizability to the entire population of children with AIS. Thus, it is proposed to develop a method of prediction of the evolution of the SIA from the diagnostic radiography to the bone maturity considering the eventual wearing of the brace using machine learning algorithms. This method is divided into two predictions: the prediction of the prescription of the brace, and the prediction of the evolution of the AIS despite brace wearing. The first prediction, based on a decision tree, obtains in learning (n = 140) an accuracy of 0.86 and in testing (n = 36) an accuracy of 0.81, a sensitivity of 0.84 and a specificity of 0.73. The second prediction, using a support vector machine approach, obtains in learning (n = 55) an accuracy of 0.80 and in test (n = 10) an accuracy of 0.80, a sensitivity of 0.33 and a specificity of 1.00. Overall, the method developed in this work proposes the prediction of the evolution of the AIS from the diagnostic radiograph with an accuracy of 0.63 by considering eight parameters from the three-dimensional reconstruction of the spine: the Cobb angle, the axial intervertebral rotation of T1-T2, T4-T5, T5-T6, T6-T7 and T9-T10, the intervertebral disc wedging of T6-T7 and the type of brace proposed in first intention. This work also proposed to redefine the progression of AIS. Usually, it is defined as a worsening of the Cobb angle of more than six degrees between the diagnosis and the bone maturity of the adolescent. Here the study of an annual evolution of AIS (degrees per year) was proposed in order to highlight the velocity of the evolution of the pathology. Although this seems to be of clinical interest, no added value was demonstrated in the prediction of the evolution of AIS despite the use of a brace. The method proposed in this work for predicting the evolution of AIS from the first diagnostic radiograph to bone maturity while considering the wearing of the brace and using machine learning techniques seems promising. No method in the literature proposed a global prediction method such as this. The addition of more patients would allow better model learning and improved results.

Date	19 Apr 2022
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Jacques A. de Guise (Supervisor) & Carlos Vázquez (Co-supervisor)