This approach to implementing a knowledge-based component attempts to adap-tively determine optimum parameter settings for groups of images on the basis of image feature vectors. The feature vector is used to group images accordingly, that in turn serve as a form of a priori knowledge for use in subsequent components. In this way components may be trained to operate on particular image groupings with different parameter settings.
In mammography, Zheng et al.  propose an adaptive computer-aided diagnosis scheme optimized on the basis of the characterization of the mammo-gram. The rule-based system proposes a difficulty index (DI). This is computed as the weighted sum of nine histogram-based features calculated from a separate training set. The computed DI score is used in conjunction with a banding scheme, based on empirically determined values corresponding to easy, moderately difficult and difficult groupings following human interpretation. An expert radiologist evaluates each mammogram and determines the group boundaries. The authors propose the use of a rule-based classification scheme such that different classification rules are independently set for the three different difficulty groups in training. On a locally defined dataset of 428 digitized mammograms (abnormal n = 220, normal n = 208), the authors report the simple adaptive scheme reduced the average number of false-positive detections from 0.85 to 0.53 per image.
Matsubara et al.  proposed the use of an image grouping scheme for digitzed mammograms. In their study, images are assigned to one of four categories based on histogram analysis of the image gray scales. Subsequent image-processing operations, such as threshold-based segmentation and region classification operate on parameters defined empirically and independently within each category. The authors use this scheme to ignore high-density mammograms. On a small dataset of 30 images, the authors report a sensitivity of 93%.
Was this article helpful?