Modélisation robuste de la texture et de la déformation de structures anatomiques par projections iconiques dans des images échographiques

Abstract

Ultrasound imaging is increasingly used because of its low cost, non-irradiating nature and portability. Indeed, the ultrasoundwaves emitted by the ultrasound probe during image acquisition are not harmful to the patient’s health. The major difficulty of ultrasound lies in the fact that the images generally have a low signal-to-noise ratio, low contrast and blurred boundaries. Most existing methods rely primarily on pixel intensity to identify anatomical structures of interest. Unfortunately, the variable quality of ultrasound images can significantly decrease their performance. In order to exploit the potential of ultrasound imaging, it is necessary to use robust methods that specifically address these challenges. The objective of this thesis is the development of robust methods for modeling the content of an image and an image sequence in the context of ultrasound imaging. To this end, an approach based on feature extraction by horizontal and vertical iconic projections of images for static and dynamic analysis of ultrasound images has been developed. Thus, this thesis resulted in three contributions. First, a method for automatic extraction of vertebral landmarks from ultrasound images in the transverse plane has been developed. Based on the detrended horizontal and vertical projections of the image, knowledge about the vertebrae is combined with the new concept of mean boundary to extract the landmarks in a fully reproducible way. The proposed approach successfully localizes the spinous process and laminae for a swine cadaver and healthy human subjects provided that the landmarks are visible in the image. Next, a global index for tissue deformation analysis was proposed. It represents a promising way to assess tissue deformation without correlation or tracking of landmarks. It is thus possible to evaluate the deformation between two images even if they are strongly decorrelated as long as the structure of interest remains visible. The results show that the index is robust and strongly correlated with the amplitude of the real deformation for simulated echocardiography (healthy and pathological cases) and ex vivo on raw meat. Finally, a method for local analysis of tissue deformation was developed as a dense, local 2D generalization of the index. For this purpose, both horizontal and vertical projections are considered while drawing on the box-counting method typically used to estimate the fractal dimension. This allows to obtain a mesh in an unsupervised way that deforms like the tissue of interest. The results obtained indicate that the method reproducibly segments the heart and accurately and robustly tracks the diaphragm over time. In conclusion, this project offers a new paradigm based on iconic image projections to robustly model the relevant information contained in ultrasound images. Although this approach was developed with the challenges of ultrasound in mind, its potential is applicable to images in general regardless of the acquisition modality.

Date	21 Mar 2022
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Catherine Laporte (Supervisor) & Farida Cheriet (Co-supervisor)