Department of Computer Architecture and Technology, University of Girona, Spain.
Acad Radiol. 2010 Jul;17(7):877-83. doi: 10.1016/j.acra.2010.04.013.
The goal of this article is to analyze and compare the performance of a developed mass computer-aided detection (CAD) system that takes breast density information into account when using manual or automatic breast density annotations in the training step. The advantages of considering this breast density information will be highlighted.
The image database used in this article is 92 mediolateral oblique (MLO) and 92 craniocaudal (CC) mammograms obtained by a full-field digital mammographic unit. All mammograms contain at least one mass. The evaluation of the experiments is performed using free receiver operating characteristic analysis for evaluating the detection performance and pixel-based receiver operating characteristic analysis for evaluating the segmentation accuracy. In addition, the performance of the automatic breast density classifier is shown using confusion matrices.
When the breast density information is not considered and at a specificity of two false positives per image, the sensitivity obtained by the CAD system is 0.747 for the CC views and 0.853 for the MLO views. Considering the breast density information, the sensitivity for CC and MLO mammograms increases to 0.800 and 0.893, respectively, using manual classification, and 0.827 and 0.907, respectively, using automatic estimation. The same trend is observed when evaluating the CAD segmentation accuracy for detected masses in terms of area under the curve values: without considering breast density, these are 0.920 +/- 0.057 and 0.917 +/- 0.072; using manual classification, 0.934 +/- 0.039 and 0.932 +/- 0.046; and using automatic estimation, 0.947 +/- 0.038 and 0.946 +/- 0.045 for CC and MLO views, respectively.
The experiments showed improved results when breast density information was taken into account. Moreover, the results obtained when using automatic breast density estimation outperformed those based on the manual annotations provided by expert radiologists. In this sense, the experiments showed that breast density information can be beneficial for CAD systems, and this information can be estimated robustly by an automatic procedure, which reduces the inter- and intra-class variability of the radiologists.
本文旨在分析和比较一个已开发的大规模计算机辅助检测（CAD）系统的性能，该系统在训练步骤中考虑了手动或自动乳腺密度标注时的乳腺密度信息。本文将重点强调考虑该乳腺密度信息的优势。
本文使用的图像数据库是由全数字化乳腺摄影设备获得的 92 张侧斜位（MLO）和 92 张头尾位（CC）乳腺 X 线片。所有乳腺 X 线片中至少包含一个肿块。通过免费的接收器操作特性分析来评估实验的检测性能，通过基于像素的接收器操作特性分析来评估分割精度。此外，还通过混淆矩阵展示了自动乳腺密度分类器的性能。
当不考虑乳腺密度信息且每幅图像有两个假阳性时，CAD 系统在 CC 视图中获得的灵敏度为 0.747，在 MLO 视图中为 0.853。考虑到乳腺密度信息，使用手动分类时，CC 和 MLO 乳腺 X 线片的灵敏度分别增加到 0.800 和 0.893，使用自动估计时分别增加到 0.827 和 0.907。在评估以曲线下面积表示的检测到的肿块的 CAD 分割精度时，也观察到了相同的趋势：不考虑乳腺密度时，这些值分别为 0.920 +/- 0.057 和 0.917 +/- 0.072；使用手动分类时，分别为 0.934 +/- 0.039 和 0.932 +/- 0.046；使用自动估计时，分别为 0.947 +/- 0.038 和 0.946 +/- 0.045。
实验结果表明，考虑乳腺密度信息可以提高结果。此外，使用自动乳腺密度估计获得的结果优于专家放射科医生提供的手动标注。在这种意义上，实验表明乳腺密度信息对 CAD 系统有益，并且可以通过稳健的自动程序来估计，从而减少了放射科医生之间和内部的变异性。
