As radiology becomes increasingly digital and picture archive and communication systems (PACS) move from research to development and practice, the quantity of digital information generated threatens to overwhelm available communication and storage media. Although these media will improve with technology, the need for efficiency will remain for inherently narrowband links such as satellites, wireless, and existing media such as twisted pair that will remain useful for many years. The expected growth in digital data as X-rays become digital will balance much of the expected gain in transmission bandwidth and local storage. Typical high-resolution digital mammograms require tens of megabytes for each image. The transfer of a collection of studies for research or education across the Internet can take hours.
Image compression can provide increases in transmission speed and in the quantity of images stored on a given disk. Lossless compression, in which an original image is perfectly
Portions reprinted, with permission, from IEEE Trans. Medical Imaging, 12(4): 727-739, Dec. 1993 and Proceedings IEEE, 82(6): 919-932, June, 1994.
recoverable from the compressed format, can be used without controversy. However, its gains are limited, ranging from a typical 2:1 compression (i.e., producing computer files of half the original size) to an optimistic 4:1. Serious compression of 10:1 or more must be lossy in that the original image cannot be recovered from the compressed format; one can only recover an approximation.
How does one evaluate the approximation? Clearly the usefulness of image compression depends critically on the quality of the processed images. Quality is an attribute with many possible definitions and interpretations, depending on the use to which the images will be put. Sometimes it is felt that for a compressed image to be considered "high quality," it should be visually indistinguishable from the original. This is sometimes referred to as "transparent quality" or "perceptually lossless" since the use of compression on the image is transparent to the viewer. Upon first consideration, this concept of visually indistinguishable would seem to be a simple definition of a perceptually important quality threshold that everyone could agree upon. This is not the case. Two images that are visually indistinguishable when seen by a certain person may be distinguishable when seen by someone else. For example, a pair of medical images viewed by lay people may appear identical, but a radiologist trained in viewing those images might detect differences. Similarly, a pair of images seen
Copyright © 2000 by Academic Press.
by the same person under certain viewing conditions may appear identical, but when seen under different conditions of ambient lighting, viewing distance, or display characteristics might be easily seen to differ. A third issue is that a compressed image might differ from the original without necessarily being worse. To hold up transparent quality as the ultimate quality goal is to ignore the possibility that certain types of computer processing, including compression, can in some cases make images more pleasing perceptually than the originals. A fourth issue is that images have different applications, and the term "high quality" may be used to denote usefulness for a specific application rather than to indicate that an image is perceptually pleasing or visually identical to an original. For all of these reasons, the measurement of image quality is a diffiult task, and only a few researchers consider quality as a binary quantity that either meets the transparent quality standard or does not. No single approach to quality measurement has gained universal acceptance. The various approaches can be categorized into the following three groups:
• Computable objective distortion measures such as squared error or signal-to-noise ratio.
• Subjective quality as measured by psychophysical tests or questionnaires with numerical ratings.
• Simulation and statistical analysis of a specific application of the images, e.g., diagnostic accuracy in medical images measured by clinical simulation and statistical analysis.
Within this latter category of evaluation methods, the methodology of receiver operating characteristic (ROC) curves has dominated historically, but a variety of other approaches have been used in which radiologists may be called upon to perform various interpretive tasks. Radiologists detect and localize the disease, make measurements of various structures, and make recommendations for patient management. The utility of a medical image can be evaluated in terms of how well it contributes to these functions.
In this chapter, we begin with a brief introduction to image compression, and to the three different sets of medical images that form the basis of our studies. We discuss signal-to-noise ratios and subjective quality ratings in the context of these data sets, as well as ROC methodology. In the next chapter, we present the clinical studies including detection, measurement, and management tasks, and in the following chapter, we discuss a number of statistical issues that arise in this sort of clinical experiment.
Was this article helpful?