Automated Quality Assurance of Plain Radiographs
by Rory L.P. McGuire

ABSTRACT

Picture Archiving and Communication Systems (PACS) are becoming
prevalent in many radiology departments. However, with such a system,
the radiographic image retake rate is still between 2.3% and 4.6%. At
this rate, tens of millions of dollars may be wasted and diagnoses
for millions of patients delayed.

The goal of this research is to explore various ways to automatically
qualify diagnostic radiographs based on criteria determined by
radiologists.  We have developed a prototype real-time radiograph
quality assurance system that could significantly reduce the overall
x-ray retake rate, lower public exposure to radiation, and implicitly
train radiography technicians by providing instant feedback on imaging
techniques.  This system could also reduce time required by the
patient, technician and radiologist, thereby reducing the total cost
of treatment.

Our system has two main parts: (1) anatomical site categorization of
an image, and (2) registration, learning, and classification of an
image for each anatomical site.  For the first part, we use mutual
information to categorize knee and chest images.  For the second part,
we focus on knee images.  We use the Hough transform, a well-known
simple edge detection algorithm; the "eigenfaces" method of
calculating the principal components of a set of images and reducing
the dimensionality of the search space; and a simple, nearest-neighbor
supervised learning and classification algorithm.  We believe our
system could serve as a springboard for similar systems to qualify
medical imaging data of other anatomical sites.