Segmentation et identification des vertèbres sur des radiographies frontales de rachis par méthodes d’apprentissage profond

Abstract

Clinical parameters extraction and 3D spine reconstruction tools from X-rays help clinicians in the diagnosis and treatment of scoliotic patients. The company EOS Imaging aims to automate such applications to better include them in clinical routine. Vertebrae segmentation and identification on X-rays represent a fundamental and crucial step in the clinical parameters and 3D spine reconstruction processes. Hence, the automatization of this relatively complex step requires particular attention. The main goal of this project is to propose an automatic method to segment and individually identify vertebrae from EOS frontal X-rays of the spine. The project leverage deep learning methods, convolutional neural networks more specifically. A database of 767 frontal X-rays from patients suffering from idiopathic adolescent scoliosis is used. Careful attention was paid to the choice of evaluation metrics so that they reflect the CNNs ability to produce clinically usable results. A comparative study of the instance segmentation CNNs DetectoRS, RetinaMask, MS R-CNN et YOLACT is conducted. DetectoRS is significantly better than all three others CNNs on the most constraining identification rate metric. The proposed method is called V-DetectoRS. The V-DetectoRS CNN is built on DetectoRS and incorporates two adjustment strategies based on the respect of the anatomic structure of the spine. A post-treatment and a penalty term added to the regression loss function of DetectoRS are added. They both ensure the respect of intervertebral distances. V-DetectoRS’s results outperform DetectoRS on every evaluation metric. Compared with the literature review methods, V-Detectors shows similar performances. V-DetectoRS reaches a mean Dice coefficient of 87.84%. Considering that a vertebra is well identified when the Dice coefficient is above 75%, 86.21% of the tested images are processed with success by V-DetectoRS. In the future, V-Detectors could be incorporated in clinical parameters and 3D spine reconstruction applications in order to participate in their automatization.

Date	23 Feb 2022
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Carlos Vázquez (Supervisor) & Jacques A. de Guise (Co-supervisor)