UCLA

Given the critical role ultrasound imaging plays in medicine, it is important to have a reliable way to measure the quality of an ultrasound image. Image quality measurements allow engineers to design ultrasound imaging systems with configurations that allow users to perform the clinical tasks and diagnoses with the best accuracy, and also allow objective comparisons to be made between different ultrasound machines. The most meaningful way to measure image quality is to conduct a study on how well humans are able to perform the clinical task they will be using the images for, but this is a resource intensive task. Therefore, in practice, ultrasound image quality is generally measured using basic mathematical metrics known as traditional image quality metrics. In this thesis, I explore how these metrics relate to human performance on a clinical task, and study whether human performance can be predicted from these metrics. To do so, I created a visual assessment which tests the ability of participants to detect artificial veins in an ultrasound image, at varying levels of noise and varying levels of undersampling (reduction in spatial resolution). I then compute the values of various traditional image quality metrics for the same ultrasound images and examine whether the trends of these traditional image quality metric values mirror the trends of human performance on our assessment, across varying levels of noise and undersampling. The results found that traditional image quality metrics had a relatively simple relationship with human performance across varying levels of noise, but a more complex relationship across varying levels of undersampling. In addition, I created regression models that were able to predict human performance from traditional image metrics for the cases studied. This work demonstrates a first step towards examining the relationship between traditional image quality metrics and task-based image quality assessments in the context of ultrasound images.