Automated image quality tool is adapted for pediatric studies

Thursday, November 30 | 11:10 a.m.-11:20 a.m. | RC613-12 | Room S102CD
In this Thursday talk, researchers from Duke University will explain how they took a tool developed for automated assessment of image quality in adult chest radiographs and adapted it for use in children.

The assessment of image quality is a key part of any quality assurance program, but this can be fraught with challenges. Traditionally, image quality has relied on just a few physical measures such as noise and contrast, while observer studies are by their very nature highly subjective, according to the group led by Dr. Ranish Khawaja, along with Dr. Gary Schooler and Ehsan Samei, PhD.

In previous work published in 2012, the Duke researchers discussed their development of an automated image quality assessment tool based on 10 perceptual attributes of adult chest radiographs. At RSNA 2017, they will discuss how they adapted the tool to work with pediatric radiographs.

The 10 attributes include factors such as lung gray levels, lung details, rib-lung contrast, and mediastinum noise. Attributes are characterized as physical quantities using an automated process along vertical lung centerlines that is highly dependent on region of interest (ROI).

When Khawaja and colleagues initially applied the adult algorithm to pediatric chest x-rays, they found poor registration of ROIs in the younger patients. So they collected 184 chest radiographs from Duke's clinical operation to develop correlation coefficients for rib width, rib interspace, and vertebral body width. These were used to adapt the ROIs for use with the image quality algorithm.

The adaptation improved the performance of the algorithm, according to the researchers. The algorithm can now be used for a variety of applications beyond image quality assessment, including protocol development, quality monitoring, and integration with dose estimates to improve benchmarks for pediatric chest radiography. The group also plans to look into whether machine learning could be used for the application.

Page 1 of 545
Next Page