Segmentation automatique de la colonne vertébrale lombaire à partir d'images à résonance magnétique par combinaison de réseau de neurones convolutifs et coupe de graphe

Abstract

Back pain, the sickness of the century as many people describe it, is a general term for a potentially serious illness and one of the most common medical problems in the world. It can occur anywhere in the spine. To identify the origin of pain and determine if treatment is needed, experts in this area rely on the analysis of medical images such as MRI and CT-scan to identify damaged areas or anomalies. The conventional radiological examination is a complicated and expensive task in precious time for the patient and the doctor. Moreover, in certain situations, identifying these anomalies with the naked eye is not always obvious, which requires the application of certain image processing techniques to guide the expert to make a good diagnosis. Among the most used techniques in this area we mention the image segmentation that delimits and identifies areas of interest. Precise and robust structural segmentation is a prerequisite for computer-assisted diagnosis and anomaly identification. It can also be used for computer-aided planning and simulated surgery. However, despite the technological inventions in this field, the approaches used for the segmentation of medical images are limited in terms of performance and require the intervention of a human expert. Recently, convolutional neural networks (CNN) have shown outstanding performance especially in the field of medical image processing surpassing existing segmentation approaches in the literature. It is in this context that this work aims to propose a new approach for the joint segmentation of the vertebrae and intervertebral discs of the lumbar spinal column based on the combination of convolutional neural networks with graph cut segmentation applied on 3D MRI images. Instead of directly applying the CNNs to obtain a final segmentation, the proposed technique uses the probability maps generated by the neural network as initialization for the graph cut method in order to refine the initial segmentation. To improve the results in the case of multilabel segmentation, we used the α −expansion algorithm which is an extension of the graph cut applied to multi-label images. The approach was quantitatively evaluated on two different databases used in the annual MICCAI competition for segmentation of vertebrae and discs. We also qualitatively evaluated our method on a new database of ten subjects that contains manual multi-label annotations of both structures ; vertebrae and discs. The experimental evaluation, based on Hausdorff distance and Dice similarity coefficient, shows that our approach performs well on all three databases.

Date	13 Aug 2019
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Ismail Ben Ayed (Supervisor) & José Dolz (Co-supervisor)