Prédiction automatique de la maturité osseuse à partir de radiographies EOS chez les patients atteints de la scoliose idiopathique de l’adolescent

Abstract

Adolescent idiopathic scoliosis (AIS) is a pathology characterized by spinal deformity. We now know that genetic factors influence the development of idiopathic scoliosis. Bone age is a good indicator of growth potential, enabling clinicians to prescribe the appropriate treatment, either surgery or the use of a corset. Traditionally, bone age was acquired by observing a 2D radiograph of the patient’s hand, using grading scales such as Tanner-Whitehouse-III (TW3) and the Sanders index. The Risser scale, however, is the most widely used for monitoring AIS patients, since it provides a radiograph, the degree of scoliotic curvature and the patient’s bone age. However, this method of assessment shows a high level of variability in reading between observers. Moreover, there is little evidence of correlation between the Risser scale and the evolution of scoliotic curvature. A recently integrated protocol at Montreal’s Sainte-Justine Hospital provides X-rays of scoliotic patients that include their hands, making it possible to observe their bone age using both traditional methods and the Risser scale. Tools for automating bone age acquisition using both traditional grading methods and more modern methods such as the Risser or Sanders scales have recently been developed, based on deep learning and machine learning models for bone age classification. The aim of this study is to develop a tool that will automatically determine the bone age of patients with adolescent idiopathic scoliosis. To this end, three models were developed to predict bone age according to the Sanders and Risser scales, using a regression model based on support vector machines. The first model was a prediction of the Risser scale based on image characteristics. The second model consisted in the prediction of the Risser and Sanders scales based on clinical characteristics. And finally, the last model consisted of a prediction of the Risser and Sanders scale based on a combination of image features. The first VGG16-SVR model demonstrated better performance for Risser-based bone age prediction with all the three databases with a maximum R2 correlation coefficient of 0.76, a mean absolute error of 0.17 and a mean squared error of 0.20. The VGG16-BoneXpert-SVR model performance for Sander index prediction was better than that of the other two models, with an R2 correlation coefficient of 0.85, a mean absolute error of 0.11 and a mean squared error of 0.15.

Date	20 Dec 2023
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Luc Duong (Supervisor)