flAdapted from source with permission.

implemented, there will be no additional work on the part of radiologists when features are collected for computer classification. Variability in radiologists' interpretation, however, will always be present in the features they report. It has long been recognized that radiologists' impressions of mammograms are not always consistent from day to day or from one radiologist to another. Studies have shown that this variability may be quite large [49-52]. Clearly, variability can affect the accuracy of computerized classification of breast lesions that is based on human perceptual features. For example, if a computer is trained by an expert radiologist and is then used by a less experienced radiologist, in order to derive the maximum benefit, the less experienced radiologist must score features in the same way as the expert. Even if the expert radiologist uses the computer aid himself, he must also score features in the same way as when he trained the computer to achieve the best computer performance. Ifthere are significant differences in the way in which features are scored, then the results of computer classification will be less accurate and the computer aid will be less effective in helping radiologists making the correct diagnoses. An example of this variability is shown in Fig. 4.

0 0

Post a comment