Apprentissage profond pour le calcul automatique de l’âge osseux à partir de radiographies postéro antérieures pour le suivi de la scoliose idiopathique de l’adolescent

Abstract

The assessment of skeletal maturity is important for the follow-up of patients with scoliosis disease, as it allows for better definition of an appropriate treatment plan for their condition. However, for its evaluation in a clinical setting, professionals often rely on a method of visual analysis of the X-ray, which is quite costly in terms of time, as clinicians are forced to abandon other tasks, probably more important, to perform the visual reading of the X-ray. As such, visual analysis also has implications in terms of accuracy, as estimates may be subjective and may vary from clinician to clinician, or even between two readings of the same X-ray by the same clinician. In this thesis, we propose an approach for the automatic assessment of bone age from EOS X rays obtained for 467 patients living with idiopathic scoliosis. After identifying the regions of interest from the X-ray, we extracted texture-based features using the Gray-Level Co-ocurrence Matrix (GLCM) technique, combined with a Support Vector Machine (SVM) model, to predict selected bone age indicators (mainly : Sanders classification, Risser index, Triradiate cartilage status and ABC scale). Furthermore, we evaluate the performance of a deep learning network (VGG16) applied directly to the regions of interest of the radiograph. The results suggest that the texture-based features extracted with GLCM are very suitable for predicting bone maturity indicators with few categories such as the Triradiate cartilage status (more than 94% of accuracy) or the ABC scale (about 82% of accuracy), while the deep learning neural network VGG16 performed better, compared to the GLCM-SVM approach, on indicators with a larger number of categories such as the Sanders classification (74% vs. 64.1% from the right hand, and 71% vs. 65.4% on the left hand) or the Risser index (84% vs. 64.3%) from the right iliac crest and 90.1% vs. 66.2% in the pelvic region. We finally made use of BoneXpert, a commercial bone age assessment tool, to extract features from the hand region in a subset of 80 radiographs. These features showed a strong correlation with the patient’s chronological age (coefficient of 0.98) and resulted in an accuracy rate of approximately 77.5% when used to predict the value of the Sanders classification associated with the radiograph. These results suggest that the deep learning approach was able to successfully explore the bone structure of the hand, following the extraction of the regions of interest from the analyzed EOS X-rays, despite the constraints encountered in terms of quality and volume of the X rays.

Date	8 Jul 2024
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Luc Duong (Supervisor) & Sylvie Ratté (Co-supervisor)