The Effects of the Computer-Aided Diagnostic System on the Diagnostic Performance

作为研究计算机辅助诊断(CAD)对诊断性能影响的基础研究，我们研究了如何将模拟结节叠加在胸部X线图像和胸部CT图像上。在这些研究的基础上，我们开发了模拟结节的数字叠加工具，并经验地推导了得到叠加结节的图像的方法。为了评估CAD输出作为“第二意见”对检测性能的影响，我们制作了5幅胸部平片图像，每幅图像包括15到19个重叠结节。对于这5幅图像，我们应用了实验CAD系统，得到了CAD输出。从这些CAD输出中，我们得到了模拟的CAD输出，其敏感度为47.6%，特异度为87.2%；该敏感度和特异度针对胸片图像上的正方形野进行了计算。两位放射科医生读取了有和没有CAD输出的5张胸片软拷贝图像，并报告了他们对胸片上正方形野内存在结节的概率判断。采用受试者工作特征(ROC)分析和Brier评分评价诊断效果。结论如下：1)无论是真阳性病例还是假阳性病例，出现结节的概率都有所增加。因此，从ROC分析来看，有CAD输出的诊断性能与没有CAD输出的诊断性能相同。2)根据Brier评分，有CAD输出的诊断性能好于无CAD输出的诊断性能，但差异无统计学意义，并利用ROC分析和Brier评分研究了重采样方法和诊断性能的评价方法。

As the baseline study for studying the effects of computer-aided diagnosis (CAD) on diagnostic performance, we have investigated how to superimpose the simulated nodule on plain chest X-ray images and on chest CT images. Based on these studies, we have developed the tool for digitally superimposing the simulated nodules, and have empirically derived the method to get the images with the superimposed nodules.To evaluate the effects of CAD output as a "second opinion" on detection performance, we have made the 5 plain chest X-ray images including from 15 to 19 superimposed nodules per one image. For these 5 images, we have applied the experimental CAD system, and we have got the CAD outputs. From these CAD outputs, we have made the simulated CAD outputs with a sensitivity of 47.6% and a specificity of 87.2% ; that sensitivity and specificity were calculated for the square fields on the chest X-ray images.Two radiologists read the 5 chest X-ray soft-copy images with the CAD outputs and without them, and reported their probability judgments regarding the presence of nodules in the square fields on the chest X-ray images. The diagnostic performance was evaluated with receiver operating characteristic (ROC) analysis and the Brier score. The conclusion was the following : 1) The probability regarding the presence of nodules increased both for the true positive cases and for the false positive cases. Therefore, from the ROC analysis, the diagnostic performance with the aid of CAD outputs were the same as that without it. 2) According to the Brier scores, the diagnostic performance with the aid of CAD outputs were better than that without it ; however, this difference was not statistically significant.Further, we have also studied the resampling method and the evaluation method of diagnostic performance by using the ROC analysis and the Brier score.