Segmentation automatique par apprentissage profond de la colonne vertébrale scoliotique à partir d'images de résonance magnétique

Abstract

This master’s project aims to have MRI assisted in planning the treatment of adolescent idiopathic scoliosis. This is a pathology that causes torsion and curvature of the vertical column with the drive of 3D complex deformations in more severe cases. MRI offers a clear visual access to both anatomical structures related to this pathology : intervertebral discs (IVDs) and vertebrae. Thoracoscopic discectomy is a minimally invasive surgical procedure of the spine used in some scoliosis treatments. Having a limited visualization of the spine during the procedure, the surgeon uses a computer navigation system that passes through an endoscopic camera that takes an image of the spine by a real patient a preoperative 3D model obtained by MRI following a 3D / 2D registration. However, the implementation of the 3D preoperative model is a segmentation of the vertebral bodies (CVs) and IVDs. This product is a simultaneous segmentation of DIVs and CVs from MRI of patients with idiopathic scoliosis. The emergence of deep learning and its successes in segmentation of images warranted by its performance and ability to generalize on various data, led us naturally to use it to solve this problem. Through the literature review, it can be seen that the simultaneous segmentation of DIVs and CVs from MRI using deep learning is not explored. Existing work focuses on one of the two structures only. Moreover, their validation on scoliotic patient data does not seem to be realized. Restricted access to medical images is a serious obstacle, especially since deep learning approaches are greedy in terms of learning data. As part of our project, we built a database from two sources. The first is composed of a large number of images of non-scoliotic subjects while the second consists of a limited number of images of scoliosis patients. We proposed a pre-treatment to clean and normalize these images in order to enhance their contrast, match the two databases and increase data through geometric strategies. The training of our model of simultaneous segmentation of DIVs and CVs goes through supervised learning in two stages. We first train a model with images of non-scoliotic subjects. Then, this model is re-trained with images of scoliosis patients. In both steps, we used a convolutional neural network (CNN) whose architecture is of encoder-decoder type. On the other hand, we introduced a feature recalibration operation named Spatial and Channel Squeeze and Excitation (scSE) to highlight the most discriminating features. Our images suffer from a class imbalance wherein the background of the image is more dominant than the DIVs and the CVs. To alleviate this problem, we have added a term based on Cohen’s Kappa coefficient to the cross-entropy function of our objective function. Our experiments show that our methodological choices are paying off. Expressing the quality of our results in terms of the Dice coefficient, we obtain in the case of CV segmentation, an average of 85% good predictions over the entire test database and an average of 78% in the case of segmentation of DIVs. We have shown that a multi-class segmentation produces better performances compared to a binary segmentation in two stages (one model for DIVs and another one for CVs). Also, the transfer of learning and the use of SWSE blocks have both revealed their interest and complementarity. The first allowed the model to become better acquainted with scoliotic patient images and thus extend the representation of DIVs and CVs to the deformations they could undergo in the context of scoliosis. The second endowed the model with a better ability to generalize, which has the direct consequence of having a broader general understanding of DIVs and CVs. The contribution we make through this project is reflected in our interest in scoliosis by validating our approach on MRI of scoliotic patients. Similar to what is proposed in the spinal segmentation literature from MRI using deep learning, our approach distinguishes by the simultaneous segmentation of CVs and DIVs, the two most important benchmarks bone in the treatment of scoliosis. In addition, the use of a strategy that takes into account the weight of characteristics through the use of the scSE block has not been explored. For future studies, we propose to extend the current architecture to include multi-modality and multi-scale information. The first information will be used to generalize the model and perform better on a larger number of MRI modalities. The second will serve to ensure an invariance with respect to the scale. To overcome the crucial problem of lack of scoliotic patient data, we propose to use a variational self-encoder in order to learn a latent representation of this scoliotic data in order to generate new images that could be used during training. Ultimately, the proposed model will allow automatic segmentation of disks and vertebral bodies that can be used in several clinical applications. Once the model is trained, real-time 3D / 2D registration of a preoperative 3D model of the spine and intraoperative radiography can be performed as part of minimally invasive spine surgery. In addition, our model will allow an automatic multimodal fusion of an MRI volume with a 3D radiographic reconstruction of the spine and the trunk surface topography, as part of the simulation of the effect of the spine surgery on the spine. external appearance of the trunk of patients.

Date	26 Apr 2019
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Catherine Laporte (Supervisor) & Farida Cheriet (Co-supervisor)