Concurrent Interactive Use of CAD for Detection of Masses in Mammograms

N. Karssemeijer, R. Hupse, M. Samulski, D. Beijerinck, G. Heeten and C. Boetes

Annual Meeting of the Radiological Society of North America 2011.

PURPOSE: To compare effectiveness of a novel interactive CAD system for detecting masses in screening mammograms to the current practice of using prompts. METHOD AND MATERIALS: We investigated a novel method of presenting CAD results to radiologists. Instead of displaying prompts after unaided inspection, information is provided on request when a suspicious region is probed using a pointer. The method is aimed at helping readers to make decisions rather than to avoid perception errors. Using an intelligent algorithm the CAD system we developed also correlates MLO and CC views, and displays corresponding CAD results in both views when only one view is probed. Suspiciousness ratings computed by CAD are displayed when regions are probed, otherwise this information remains hidden. In an observer study we compared effectiveness of this interactive system to traditional prompting. Five certified mammographers and three residents participated in the study. After a short training phase, an enriched series of 200 FFDM screening cases with priors was read two times, once with interactive CAD and once with CAD prompts. In the session with CAD prompts findings were recorded before and after activating CAD, to allow assessment of unaided performance. The study series included 63 screen detected cancers, 17 missed cancers that were retrospectively visible, 20 false positives, and 100 normals. Microcalcification cases were excluded. Readers reported localized findings with malignancy ratings scored on a continuous scale. Correct cancer localization fraction was computed per reader as a function of the false positive fraction. RESULTS: Readers prefered the novel CAD system. Because they were not distracted by irrelevant false positive CAD prompts they had more confidence in the computer aid. A significant improvement in detection performance was obtained with interactive CAD (p < 0.012) while CAD prompts had no effect on detection performance. Mean sensitivity in an interval of negative recalls less than 15% was 49.8% when reading without CAD, 49.7% with CAD prompting, and 57.7% with interactive use of CAD. CONCLUSION: Detection performance for mammographic masses increased significantly with concurrent interactive use of CAD. while presenting CAD prompts had no effect. CLINICAL RELEVANCE/APPLICATION: Mammography CAD for masses may become more effective when it is used for decision making. This may lead to earlier detection of invasive cancers.