Although commercial computer-aided detection (CAD) products have been used in a large number of medical institutions to assist radiologists in screening mammography, radiologists often have limited confidence in CAD cueing results for masses due to the relatively low sensitivity and reproducibility, as well as the higher false-positive rate, than that compared with the performance for microcalcification detection and characterization. There is no general agreement on whether and how a CAD system can best help improve radiologists' diagnostic performance. To address these issues, we propose to develop and evaluate a unique Interactive Computer-Aided Detection (ICAD) system for mammography. In addition to providing initial cues on the processed images as would current systems, radiologists can interact with the ICAD system in several ways. Observers can select any regions (cued or not cued) on the mammogram and query ICAD. Upon receiving a request from the observer, ICAD will extract the region, compute a feature vector, compare the vector with a large number of regions in a reference library, and generate a classification score using a weighted k-nearest neighbor algorithm. In addition, upon request, the system will provide the results of a scheme specifically optimized on "the latest prior images" to enable observers to assess "early signs" for abnormalities that may develop and later (on subsequent examinations) prove to be malignant. To test such a hypothesis, an observer performance study will also be carried out in this project to compare the performance difference when six radiologists use a CAD system operating at the performance level of leading commercial products and our newly proposed ICAD system.

R01CA101733

2004-09-01

2010-06-30