Affecting observer behavior and performance with a new approach to CAD

点击翻译按钮获取中文摘要

 DESCRIPTION (provided by applicant): Computer Aided Detection (CAD) tools are routinely used during the interpretation of mammograms. CAD systems are very good at detecting and highlighting (marking) clusters of micro-calcifications, regardless if a radiologist uses CAD as originally intended, namely as a second reader, or as a marker for what to look at, or not, after initial review. When it comes to identifying soft tissue abnormalities, such as masses, asymmetries, or distortions, CAD systems do not perform as well. Consequently, to maintain some reasonable level of sensitivity, many false positive marks need to be identified on the images. On average, a little less than one mark per image is highlighted by these systems. Hence, for a four view examination, it is very common that more than one mark identifying possible soft tissue abnormalities will be provided. As a result, radiologists discard a large number of CAD "suspected" regions and tend to have little confidence in these marks. It has been well documented that the impact on observer performance is minimal, if any, in detecting additional cancers depicted as soft tissue abnormalities. However, the distribution of these marks is non-uniform and depends on both the breast density and tissue structure. As these marks may range from none to as many as twenty marks per case, companies basically artificially limit the total number of marks provided per examination by setting an image based or a case based threshold, and all marks with lower CAD output scores are automatically discarded (not shown to the interpreter). We know from many studies that false negative interpretation of cancers depicted as soft tissue abnormalities are common. As important, we know that in retrospective reviews of missed and detected cancer cases a large fraction of the cancers are actually depicted on prior mammograms one year (or more) earlier than the examination resulting in the "detection" of the cancer. We also know that quite frequently cases that are hard for the observer to interpret are also "more difficult" for the CAD to analyze in terms of performance, namely, they tend to have a larger number of possible marks and frequently higher CAD output scores. There are no studies on how distributions of the number of marks in false negative (or interval) cases and cases depicting abnormalities but detected at a scheduled subsequent examination differ/compare with that of marks in examinations leading to a cancer "detection" (we combine the two groups, of interval cancers and visible but detected at a subsequent scheduled examination, and term them as "missed" cancer). We propose to use a modified CAD system to highlight specific cases rather than "suspected" regions and test in an observer study whether a totally different CAD that provides "warnings" when a case is likely to belong to the type in which there is higher likelihood to miss cancers affects observers' behaviors and performances. Our study design will enable us to compare observer performance "without" CAD results with the use of a conventional, as well as a new and innovative, CAD system.