Li Hui, Giger Maryellen L, Olopade Olufunmilayo I, Lan Li
Department of Radiology, The University of Chicago, 5841 S. Maryland Avenue, Chicago, IL 60637, USA.
Acad Radiol. 2007 May;14(5):513-21. doi: 10.1016/j.acra.2007.02.003.
To evaluate fractal-based computerized image analyses of mammographic parenchymal patterns in the task of differentiating between women at high risk and women at low risk for developing breast cancer.
The fractal-based texture analyses are based on a box-counting method and a Minkowski dimension, and were performed within the parenchymal regions of normal mammograms. Four approaches were evaluated: 1) a conventional box-counting method, 2) a modified box-counting technique using linear discriminant analysis (LDA), 3) a global Minkowski dimension, and 4) a modified Minkowski technique using LDA. These fractal based texture features were extracted from regions of interest to assess the mammographic parenchymal patterns of the images. Receiver operating characteristic analysis was used to evaluate the performance of these features in the task of differentiating between the two groups of women.
Receiver operating characteristic analysis yielded an A(z) value of 0.74 based on the conventional box-counting technique and an A(z) value of 0.84 based on the global Minkowski dimension in the task of distinguishing between the two groups. By using LDA to assess the characteristics of mammograms, A(z) values of 0.90 and 0.93 were obtained in differentiating the two groups, for the modified box-counting and Minkowski techniques, respectively. Statistically significant improvement was achieved (P < .05) with the new techniques compared to the conventional fractal analysis methods. A simulation study, which used the slope and intercept extracted from the least square fit of the experimental data with the LDA approaches, yielded A(z) values similar to those obtained with the conventional approaches in the task of differentiating between the two groups.
The proposed LDA approach improved significantly the separation between the two groups based on experimental data. Because this approach was used as a linear classifier rather than as a regression function, it combined the fractal analysis with the knowledge of the high- and low-risk patterns, and thus better characterized the multifractal nature of the parenchymal patterns. We believe that the proposed analyses based on the LDA technique to characterize mammographic parenchymal patterns may potentially yield radiographic markers for assessing breast cancer risk.
评估基于分形的计算机图像分析在鉴别乳腺癌高风险女性和低风险女性的乳腺实质模式任务中的应用。
基于分形的纹理分析基于盒计数法和闵可夫斯基维数，在正常乳腺X线片的实质区域内进行。评估了四种方法：1）传统盒计数法；2）使用线性判别分析（LDA）的改进盒计数技术；3）全局闵可夫斯基维数；4）使用LDA的改进闵可夫斯基技术。从感兴趣区域提取这些基于分形的纹理特征，以评估图像的乳腺实质模式。采用受试者操作特征分析来评估这些特征在区分两组女性任务中的性能。
在区分两组的任务中，基于传统盒计数技术的受试者操作特征分析得出的A(z)值为0.74，基于全局闵可夫斯基维数的A(z)值为0.84。通过使用LDA评估乳腺X线片的特征，改进盒计数法和闵可夫斯基技术在区分两组时分别获得了0.90和0.93的A(z)值。与传统分形分析方法相比，新技术取得了具有统计学意义的显著改进（P < 0.05）。一项模拟研究使用从LDA方法的实验数据最小二乘拟合中提取的斜率和截距，在区分两组的任务中得出的A(z)值与传统方法获得的值相似。
基于实验数据，所提出的LDA方法显著改善了两组之间的区分。由于该方法用作线性分类器而非回归函数，它将分形分析与高风险和低风险模式的知识相结合，从而更好地表征了实质模式的多重分形性质。我们认为，所提出的基于LDA技术表征乳腺实质模式的分析可能会产生用于评估乳腺癌风险的影像学标志物。
