3D Medical Ultrasound Deconvolution Project

Project Overview

Medical ultrasound machines work by sending a sequence of sound beams into the body which are reflected back from organs and tissue boundaries. Very high frequency sound (ultrasound) travels in straight lines, and this tells us the line on which a reflector is located. If we know how fast the sound travels, we can calculate the depth of the reflector along this line from the time it takes for the reflected echo to return. Thus we can build up a picture of what is going on inside the body.

The problem with this sort of imaging is that we cannot make the sound beams thin enough, and as a result, each bit of our output data is derived from the interaction of reflectors in a small volume around the point we are trying to scan. This leads to a particular type of blurring in our output images called speckle.

The goal of this project is to reduce this problem and produce clearer ultrasound images with better resolution. We will do this by first scanning known shapes to work out the three-dimensional volume in which the reflectors interact to create the speckle. As the beam-width varies with distance from the probe, the size and shape of this volume changes with depth and we will have to measure it several times and model the variation.

We will then use our model of this volume at each point in the ultrasound image, together with the reflected raw ultrasound signal, to estimate the reflectors in the body using a technique called deconvolution. This reflector information can then be used to create clearer images with less speckle. We will evaluate the usefulness of our new images at Addenbrooke's Hospital, particularly focussing on scans of thyroids, breasts, arteries, and the shoulder.

Doctors frequently want to work out the the size and shape of a particular structure in the body. For example it is necessary to monitor cancerous growths to check that they are getting smaller in response to treatment. To do this, doctors have to identify the scan data relating to the structure of interest. This process is called segmentation.

It is very hard to design reliable automatic algorithms using a normal ultrasound image. Most doctors end up just drawing round the object of interest using a mouse or pointer. One of the reasons for this is that ultrasound images are created using signal processing techniques that discard a lot of the information in the original signal and are thus hard to relate to the physical properties of the tissue being scanned. When we calculate the reflector data we will preserve the full information content of the signal. We therefore hope to be able to devise more reliable segmentation algorithms to work on this rich representation. We will compare our new algorithms with segmentations performed by doctors to evaluate their performance.

All this work is only possible if we work with three-dimensional ultrasound data. Our group has a strong track record in this area, and recently developed the highest definition freehand three-dimensional ultrasound system on record. Through this project we will open up a new era of clearer, higher-resolution, cheap, safe, ultrasound imaging.