While many investigators have made great progress in developing methods for computer-aided detection and diagnosis (CAD) of lesions, current human interfaces for communicating the computer output to the user are inadequate. Intelligent workstations that aid radiologists in diagnosing cancer utilize an estimate of a lesion's probability of malignancy, usually obtained by training a classifier on an independent database. These estimates of the probability of malignancy are dependent on the prevalence of cancer in the training database, which most often does not correspond to the prevalence of cancer in the population from which the user has experience, e.g., the population seen in the user's medical practice. Thus, the user often has difficulty interpreting the computer-estimated probability of malignancy. This proposal aims to extend our intelligent workstation to include a transformation of the computer output that is either reader specific or radiology-practice specific.
The specific aims of the study are (1) collect radiologists'rating data on a database of clinical mammograms and sonograms in terms of their assessment of the probability of malignancy, (2) develop models with which to transform computer output to those that would """"""""match"""""""" the internal parameters of the reader, (3) compare the models using both the computer and human data across the two modalities, (4) use the results of the transformation to modify computer output in our design of clinically useful interfaces, and (5) perform the first test of CAD using enhanced intelligent interfaces customized to the particular radiologist or the particular radiology practice. The proposed study is highly relevant in that such enhanced human computer interfaces are expected to improve and expedite the use of computer aids in breast cancer imaging and interpretation.
