Computer-aided diagnosis (CADx) schemes are being developed to help radiologists distinguish benign from malignant lesions. 1 approach is to recall from a library of cases with known pathology, lesions that are similar to the image in question. The difficulty of this approach is determining which lesions from the library are similar to the unknown case. In current approaches, the recalled lesions do not always look similar to the unknown case. We propose to develop a new method based on how feature values differ from the mean value of the population. That is, the most unusual features of a lesion are the prominent visual features and should be matched to find a similar lesion. Our hypothesis is that this new method of determining similarity will produce a set of lesions that are more similar to an unknown lesion than sets produced using existing methods. Specifically, we will: 1. Assemble a database of images with clustered calcifications. 2. Develop the similarity method based on difference from the mean feature value; and 3. Evaluate the performance of the method in a reader study against existing similarity methods. If we are successful, we will have improved the utility of CADx systems in which similar images are used to assist radiologists in distinguishing benign from malignant lesions. This type of system has the potential to reduce the number of biopsies of benign lesions and improve the sensitivity of the imaging modality. A reduction in the number of """"""""unnecessary"""""""" biopsies will reduce cost, and physical and mental trauma to patients. ? ?
