Abstract

Ultrasound elastography involves measuring the mechanical properties of tissue, and has many applications in diagnostics and intervention. Ultrasound elastography techniques mainly target obtaining strain images from raw Radio-Frequency (RF) echo field produced by ultrasound machine without adding any hardware. A common step in different elastography methods is imaging the tissue while it undergoes deformation and estimating the displacement field from the images. A popular next step is to estimate tissue strain, which gives clues into the underlying tissue elasticity modulus. To estimate the strain, one should compute the gradient of the displacement image, which amplifies the noise. The noise is commonly minimized by least square estimation of the gradient from multiple displacement measurements, which reduces the noise by sacrificing image resolution.
The first part of this thesis propose a new method which adaptively adjusts the level and orientation of the smoothing strain images using two different mechanisms. First, the precision of the displacement field decreases significantly in the regions with high signal decorrelation, which requires increasing the smoothness. Second, smoothing the strain field at the boundaries between different tissue types blurs the edges, which can render small targets invisible. To minimize blurring and noise, we perform anisotropic smoothing and perform smoothing parallel to the direction of the edges. The first mechanism ensures that textures/variations in the strain image reflect underlying tissue properties and are not caused by errors in the displacement estimation. The second mechanism keeps the edges between different tissue structures sharp while minimizing the noise.
The second part of this thesis introduces a 2D strain imaging technique called SHORTCUT (meSHing Of gRadienT in DP for direCt Ultrasound elasTography) based on minimizing a cost function. The cost function incorporates similarity of echo amplitudes and tissue continuity. The proposed technique is fast, robust and accurate and it directly produces the strain images from RF data using a novel dynamic programming (DP) configuration. Unlike the standard DP algorithm which discretizes the decision space (displacement field) and search in the space of piecewise constant functions, the proposed DP discretizes the gradient of the decision space (strain field) and search the space of continuous piecewise linear functions. Eliminating the displacement differentiation block and performing a global search instead of local search which exist in all of the available strain estimation techniques result in substantial improvement in SNR, CNR and accuracy of the estimations. The effectiveness of the proposed methods is investigated through simulation data, phantom experiments, and in vivo patient data.