DETECTION AND LOCALIZATION IN MAMMOGRAPHY
It is widely believed that computer-aided detection (CAD) schemes will eventually provide useful prescreening tools or "second opinions" in radiology, to improve efficiency and accuracy in the detection of breast cancer-particularly subtle cancers at early stages in clinical screening environments. Existing CAD schemes typically provide their outputs by marking (cueing) suspicious locations on images for the reader's attention. Recent evidence, however, suggests that such cues can have detrimental effects when prompts are presented before readers have an opportunity to form unbiased impressions about the case findings. Current CAD schemes treat single images in isolation, whereas clinical mammograms obtain two separate views of both breasts (4 images). This project uses prospective interpretations of clinical mammograms to study the interface between a CAD system and the physician. Reader-performance studies will measure and compare radiologists' accuracy in detecting and localizing breast cancers, both with and without CAD cues, applying a concurrent analysis of performance that improves the statistical estimates of detection accuracy. Study 1 will use single-view mammograms to compare two different protocols for presenting CAD cues to readers, either prior to the case interpretation or only after an initial umprompted reading of each case. This study will investigate four different simulated CAD schemes that vary both the fractions of actual lesions prompted and the cueing rates for false-positive prompts. Study 2 will evaluate CAD prompting for prospective interpretations of complete, two-view mammograms (4-images). It will also investigate augmented CAD cues that code information about correlations between the two ipsilateral projections.