Tissue deformation analysis using ultrasound images

Abstract

Tissue deformation analysis aims at studying the elastic properties of soft tissues. Such properties provide unique information for clinical diagnosis. In order to measure the elastic properties of tissues, a source of mechanical vibration is needed to induce tissue motion and an imaging modality is used to record images of tissues. Motion tracking methods are then applied to infer the tissue motion / deformation. In this thesis, ultrasound image based tissue deformation analysis is studied. A difficult problem called feature-motion decorrelation, which restricts the effectiveness of the tracking methods, is solved using a coupled filtering method and an affine warping method. The performance of both methods is evaluated with a comparison with the direct correlation method, which does not facilitate any compensation of decorrelation. Results show that the coupled filtering method and the affine warping method are able to achieve a robust deformation estimation even when tissues undergo extremely large deformation, while the direct correlation method fails when tissue deformation is large. The performance of the affine warping method is also evaluated using B-mode (BM) images instead of radio-frequency (RF) signals, and the coupled filtering method and the affine warping method are compared with one of the state-of-the-art methods in ultrasound based tissue deformation analysis. Tracking methods that only model translation, compression and expansion as in most of the state-of-the-art methods will fail in a typical elastography setting with multiple stiffness regions in tissues. Ultrasound based image registration, a problem similar to the problem of ultrasound based motion analysis, is also studied. In particular, a groupwise registration method is proposed for 3-D echocardiography. The proposed method aligns 3-D ultrasound image volumes of the heart from different view angles. After the alignment, image information from multiple sequences can be incorporated for further processing. Sparse and low rank modeling of ultrasound image volumes is used. The alignment of image volumes is facilitated by finding a minimum number of linear bases to approximate all volumes. The linearly correlated approximations serve as the fused volumes and in the fused volumes speckle patterns are reduced. The fused volumes serve as the reference for the registration and in the proposed method no reference image volumes need to be chosen. By using the proposed method, missing structures in single-view volumes can be recovered by modeling them as sparse outliers.

Date of Award	2014
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology