SPATIALLY-VARYING BLIND DECONVOLUTION IN ULTRASOUND LOCALIZATION MICROSCOPY

Jiang, Bowen

Permalink

https://hdl.handle.net/2142/124949 Copy

Description

Title

SPATIALLY-VARYING BLIND DECONVOLUTION IN ULTRASOUND LOCALIZATION MICROSCOPY

Author(s)

Jiang, Bowen

Issue Date

2021

Keyword(s)

image processing; blind deconvolution; sparse recovery; ultrasound localization microscopy; deep learning.

Abstract

Ultrasound Localization Microscopy (ULM) has been widely used in microvascular imaging. ULM uses ultrasonic waves reflected from micro-bubbles injected as an intervascular contrast agent to localize the micro-bubbles at a resolution exceeding the diffraction limit. However, the observed ULM images formed by beam forming are blurred by unknown and spatially-varying point spread functions (PSF) due to sound propagation in tissue with inhomogeneous and imprecisely known sound velocity, and near-field propagation effects. To overcome this blurring, blind deconvolution is necessary. In this thesis, we analyze two different algorithms to solve the ULM image deconvolution problem: a recently proposed Multichannel Sparse Blind Deconvolution (MSBD) algorithm, and a fully convolutional network (FCN). Specifically, given multiple frames of blurred observations, the MSBD algorithm is an optimization that aims to find a spatially-invariant deconvolution f ilter, such that applying it to the blurred observations results in images that are as sparse as possible using a certain sparsity metric. To use the MSBD algorithm in the spatially-varying case, we apply it in a sliding-window fashion to subimages, assuming that the PSF within each window is spatially invariant. Although the estimates of the sparse images produced by the MSBD algorithm are not satisfactory by themselves, the MSBD estimate of the approximate PSF for each window is accurate enough to enable satisfactory non-blind deconvolution using the Hogbom CLEAN algorithm. Our second, FCN-based algorithm, also operates on sliding windows, but in the inference stage it processes one window from a frame at a time. The FCN is trained end to end in a supervised manner using the cross-entropy loss, and is able to learn to deconvolve blindly different spatially-varying PSFs using a training dataset consisting of known ground-truth images and their blurred versions. A limitation of this approach may be limited generalization to scenarios substantially different from those in the training set.

Type of Resource

text

Language

eng

Owning Collections