| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Letters to the Editor |
930 East Emerald Avenue, Knoxville, TN 37917. e-mail: amde-1@comcast.net
Editor:
I read with great interest the article by Yoshida and colleagues in the February 2002 issue of Radiology (1). They demonstrated improved sensitivity for detection of polyps by using computer-aided diagnosis (CAD). Computers have also demonstrated diagnostic usefulness in other areas of radiology, including mammography and nuclear medicine lung scanning, among others. Although still in their infancy, CAD and artificial neural networks continue to progress.
While I take some measure of pride in seeing such cutting-edge work performed at my old alma mater in Chicago, some disturbing questions come to mind concerning CAD. In particular, what is the logical end point of CAD?
It is obvious that computers are evolving rapidly in a very short time frame (Moore’s law). As they exponentially become more powerful, software applications continually improve. Granted, visual patterns are challenging for a computer, but it is a certainty that over time the ability of the computer to recognize patterns (which is basically what radiologists do when reading images) will continue to improve. While the high false-positive rates currently seen in CAD require the expertise of a radiologist to separate the significant findings from the insignificant ones, with time the high false-positive rate will likely decrease as the programs improve.
There is no reason to assume that CAD has to be a tool solely limited to radiologists. It could be used by health maintenance organizations, hospitals, family physicians, nonradiology specialists, and the like. Several scenarios are conceivable. For example, why could not a group of gastroenterologists use CAD to read computed tomographic (CT) colonograms? As mammography computer programs improve, assuming a high true-negative rate and a reasonable false-positive rate, they could be used by gynecologists, at their own mammography centers, to separate the many negative mammograms from the few questionably positive ones. The few questioned cases could then be reviewed by a contracted radiology group (thereby ending the shortage of mammographers). There is no reason that CAD cannot someday be expanded to all imaging aspects of radiology, including magnetic resonance imaging. Theoretically, a group of general practitioners could set up an imaging center and have all the cases reviewed with CAD. With the aid of a computer, the difference in interpretation skill between general practitioners and that of radiologists could become negligible especially for straightforward cases. Again, difficult cases could be contracted out to radiologists.
Human beings will always be subject to error. As radiology continues to expand, keeping up with all of it becomes all but impossible, and our knowledge base gets spread thinner and thinner over each modality. The computer programs alternatively will only get better with time.
Will computers eventually outperform radiologists (similar to the way chess programs now regularly defeat human chess masters), making the radiologist superfluous?
REFERENCES
Department of Radiology, University of Chicago, 5841 South Maryland Avenue, MC2026, Chicago, IL 60637. e-mail: yoshida@uchicago.edu
We appreciate the comments made by Dr Evancho and his interest in our study (1). We would also like to acknowledge his observation that CAD will be useful in various areas of diagnostic radiology, including the detection of polyps at CT colonography.
As Dr Evancho mentioned, the performance of computers will improve with time, and thus the ability of the computer to recognize patterns characteristic of lesions is also likely to increase. More and more sophisticated computer-vision algorithms will be developed, and algorithms that are effective but were too slow to run on a workstation several years ago will run rapidly enough on the high-speed workstations of today. Computers make diagnostic errors, as do human beings. However, together they can improve the diagnostic performance, as demonstrated in clinical studies in which CAD has been shown to reduce human errors and, as a result, to improve the diagnostic accuracy of radiologists (2).
Dr Evancho raised an intriguing point of using CAD as a tool for narrowing the interpretation skill between expert radiologists and the nonexperts. Indeed, findings of observer studies show that such a trend is observed in CAD for mammography and chest radiography (3,4). In this sense, CAD can be regarded as a means for translation of the interpretation skills of experts to nonexperts. Aside from ethical issues, it would be possible in the future for CAD to be used to separate negative cases even before physicians read the cases. However, it should be noted that such a separation is likely to come with an increased number of missed abnormalities. Because the diagnostic performance of a CAD scheme is represented by a receiver operating characteristic curve, there is always a trade-off between sensitivity and specificity. Thus, when a CAD scheme is set to yield a high true-negative rate (ie, a high specificity), the price we need to pay is a high false-negative rate (ie, a low sensitivity) (5). Researchers in the field of CAD are working intensively to develop CAD schemes that yield both high sensitivity and high specificity, so that a radiologist with computer aid would eventually miss no malignancies and call no benign lesions abnormal. This will continue to be a main research goal in CAD.
Although CAD is a promising technique, there is a long way for CAD to go to achieve this goal. In particular, unlike mammography where years of clinical experience preceded CAD, CAD in CT colonography is still in its infancy. New pitfalls for human observers and computers are being discovered. Flat or infiltrative lesions are difficult to recognize with a computer. Extracolonic findings are a potentially important aspect of CT colonography, which requires interpretation by radiologists.
The ultimate goal of CAD is to mimic the role of a radiologist in the sense that a CAD scheme will output a diagnosis by means of interpretation of image data and clinical information. The potential of CAD will be most important in decision making in complex situations such as differential diagnosis, where the performance of an artificial neural network could exceed that of a human observer (6). In the future, it is expected that the interpretation of all types of medical images will make use of some form of CAD.
In large clinical studies conducted to date (2), CAD has been shown to be beneficial in relatively simple focused tasks such as the detection of microcalcifications, and CAD remains a diagnostic tool that leaves the final diagnosis to radiologists. Although CAD has the potential to replace some parts of the radiologists’ diagnostic tasks in the future, this is likely to happen in areas where rendering a diagnosis is relatively well defined but tedious. In these areas, computers can be quick, objective, and consistent and thus can help radiologists do their jobs faster and better than they already do, that is, a single radiologist could read more cases more efficiently. This would be of particular benefit in the era where the ability of modern imaging devices such as multi–detector row CT is overwhelming the ability of radiologists to interpret each image carefully.
As the benefits of CAD are established, it will become more difficult to justify not using CAD, just as it would be difficult for a radiologist to justify not using a magnifying glass for reading mammographic images. Thus, it would be possible in the future that a set of specialized CAD systems will be operated by nonradiologists in the diagnosis of a large number of mostly normal cases, for example, for screening purposes. However, we see this as an opportunity for radiologists to expand their sphere of influence by placing these CAD systems under their control, rather than losing procedures irretrievably to other specialists. CAD will provide a powerful diagnostic tool that should allow radiologists to get back on top of their specialty, rather than making them superfluous.
REFERENCES
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
