Health Devices. 2003 Apr;32(4):141-64.
While mammography is regarded as the best means available to screen for breast cancer, reading mammograms is a tedious, error-prone task. Given the repetitiveness of the process and the fact that less than 1% of mammograms in the average screening population contain cancer, it's no wonder that a significant number of breast cancers--about 28%--are missed by radiologists. The fact that human error is such a significant obstacle makes mammography screening an ideal application for computer-aided detection (CAD) systems. CAD systems serve as a "second pair of eyes" to ensure that radiologists don't miss a suspect area on an image. They analyze patterns on a digitized mammographic image, identify regions that may contain an abnormality indicating cancer, and mark these regions. The marks are then inspected and classified by a radiologist. But CAD systems provide no diagnosis of any kind--it's up to the radiologist to analyze the marked area and decide if it shows cancer. In this Evaluation, we describe the challenges posed by screening mammography, the operating principles and overall efficacy of CAD systems, and the characteristics to consider when purchasing a system. We also compare the performance of two commercially available systems, iCAD's MammoReader and R2's ImageChecker. Because the two systems offer comparable sensitivity, our judgments are based on other performance characteristics, including their ease of use, the number of false marks they produce, the degree to which they can integrate with hospital information systems, and their processing speed.
虽然乳房X光检查被认为是筛查乳腺癌的最佳可用方法，但解读乳房X光片是一项繁琐且容易出错的任务。鉴于该过程的重复性以及平均筛查人群中不到1%的乳房X光片含有癌症这一事实，难怪有相当数量（约28%）的乳腺癌被放射科医生漏诊。人为误差是如此重大的障碍，这使得乳房X光检查筛查成为计算机辅助检测（CAD）系统的理想应用。CAD系统充当“第二双眼睛”，以确保放射科医生不会错过图像上的可疑区域。它们分析数字化乳房X光图像上的模式，识别可能包含表明癌症的异常区域，并标记这些区域。然后由放射科医生检查并分类这些标记。但CAD系统不提供任何类型的诊断——由放射科医生分析标记区域并确定其是否显示癌症。在本评估中，我们描述了乳房X光检查筛查带来的挑战、CAD系统的工作原理和总体功效，以及购买系统时要考虑的特征。我们还比较了两款商用系统——iCAD的MammoReader和R2的ImageChecker的性能。由于这两款系统具有相当的灵敏度，我们的判断基于其他性能特征，包括它们的易用性、产生的错误标记数量、与医院信息系统集成的程度以及它们的处理速度。
