M.Sc Thesis


M.Sc StudentIvry Mark Noy
SubjectStructural Enhancement and Despeckling for the Production
of Panoramic Ultrasound Images using Deep Neural
Networks and Computer Vision
DepartmentDepartment of Biomedical Engineering
Supervisor PROFESSOR EMERITUS Dan Adam
Full Thesis textFull thesis text - English Version


Abstract

The World Health Organization states that low back pains (LBPs) are one of the dominant factors that restrict motion and working abilities, leading to economic and social burden. Many methodologies relieve LBPs in patients, including mainly injection-based treatments. Practically, precise injections are challenging, which renders imaging an essential guiding tool for physicians. Common imaging approaches include either a computed tomography (CT) or an X-ray scanning. While the former restricts offline analysis, the latter exposes physicians to radiation. A radiation-free, real-time imaging approach solution is offered by Ultrasound (US) imaging. US images are produced by reflections of high-frequency sound waves passing through body structures. The great importance of medical US imaging entails high significance in enhancing US images quality, contrast, and field of view (FOV). Compared to CT or X-ray scans, US images are characterized by degraded quality and low contrast due to speckles and noise interference. These artefacts might conceal the desired tissues and deteriorate detection of meaningful patterns. Also, FOV of US images is limited by several factors; the prob used for image acquisition, the medium structure via which ultrasound waves pass, and the volume and position of the organs. Enhancement of US images is achieved by two main approaches: despeckling and intrinsic objects enhancement and increasing the image FOV. To obtain image enhancement while reducing speckles, we combine deep neural networks (NN) and diffusion maps (DM) - a non-linear dimensionality reduction technique that preserves intrinsic local structures via manifold learning. We consider a sequence of US images, calculate its corresponding DM embedding, and use it to reconstruct the image. This process is done via a deep encoder-decoder NN architecture, which encodes the original US image into its DM coordinates and decodes it back into an image. This constrains the image reconstruction to contain meaningful intrinsic structures, rather than interferences. The second part of this research deals with the extensive problem of producing accurate wide-frame visualization of the underlying tissue structures. The fundamental components of this framework are US frame registration and stitching, where the main task is to tackle the challenging conditions caused by non-rigid deformations of the image structures. These are caused by both user-operation of the US imaging device, and by the movement of tissues. These ingredients allow establishment of a wide FOV image, which demonstrates an enhanced view of the organ under investigation. For improved image resolution and de-speckling, we exploit the mutual information of overlapping regions in acquired images. This notion is applicable, since a set of images, taken from a range of angles, holds a potential for super-resolved image reconstruction of thin features. In this work we use sequences of US images taken from health people as a preliminary step to improve the guidance of the physicians during the RFA procedure. Performance assessment of the mentioned algorithms was measured using several quantitative metrics including SSIM, noise reduction estimation, and SNR. Our method outperformed competing methods in terms of both structure preservation and enhancement, and noise and speckles suppression.