THe Effect of the Computer-Aided Diagnosis on the Diagnostic Performance

作为分析计算机辅助诊断(CAD)对放射科医生在筛查乳房X光检查中诊断性能的影响的基础研究，我们调查了观察者之间在判断乳房X光检查结果时的一致性。然后，为了分析计算机辅助诊断在胸部螺旋CT筛查中的作用，我们评估了放射科医生在螺旋CT上对模拟肺结节的检测能力，并通过图像读取研究，评估了计算机辅助诊断输出作为“第二意见”对放射科医生检测诊断能力的影响。在这里，我们研究了不同灵敏度和特异度的模拟CAD的效果。6位经验较少的放射科医生阅读了100幅信号脉冲噪声图像和100幅纯噪声图像，这些图像由计算机生成并随机显示在CRT上。他们报告了他们对背景噪声中存在线条的概率判断。采用受试者工作特征(ROC)分析评价放射科医师的工作表现，以双正常ROC曲线下面积(Az)作为评价指标。通过对D.D.Dorfman等人提出的26种图像读取实验的伪值方差分析(ANOVA)，分析了26种图像读取实验的Az差异。结论如下：1)在精度较高的情况下，借助CAD系统的诊断性能要好于没有CAD系统的情况。2)CAD输出的整体准确性是影响放射科医师诊断效果的最重要因素。在这里，在CAD输出的总体精度相同的情况下，CAD输出的敏感性高于特异性时，放射科医生的表现会更好。

As the baseline study to analyze the effects of the computer-aided diagnosis (CAD) on the diagnostic performance of radiologists in screening mammography, we have investigated the interobserver agreement in juding the findings of mammography. Then, to analyze the effects of CAD in screening thoracic helical CT,we have evaluated the radiologists' performance in detection of simulated lung nodules on helical CT.Further, we have evaluated the effects of CAD output as a "second opinion" on radiologists' performance in the detection diagnosis, by the image-reading study. Here, we have studied the effect of simulated CAD's with various sensitivities and specificities. Six radiologists with less experience read 100 signal pulse noise images and 100 noise only images that were produced by the computer and were randomly displayd on CRT.They reported their probability judgments regarding the presence of lines in the background noise. The radiologists' performance was evaluated with receiver operating characteristic (ROC) analysis, and A _z (the area under the binormal ROC curve) was used as an index of performance. The difference among A_z OF 26 kinds of image-reading experiments was analyzed by the analysis of variance of pseudovalues (ANOVA) computed by the jackknife method ; this ANOVA method was proposed by D.D.Dorfman et al.The conclusion is the following : 1) The diagnostic performance with the aid of CAD systems with the rather good accuracy is better than that without the aid of CAD systems. 2) The overall accuracy of CAD outputs is the most effective factor to affect radiologists' performance in the detection diagnosis. Here, in the cases in which the overall accuracy of CAD outputs are the same, radiologists' performance will be better in the higher sensitivity of CAD outputs than in the higher specificity.