Pattern Recognition Lab, Friedrich-Alexander University Erlangen-Nuremberg, Erlangen, Germany,
Int J Comput Assist Radiol Surg. 2014 May;9(3):379-86. doi: 10.1007/s11548-013-0969-5. Epub 2013 Dec 11.
Several cell detection approaches which deal with bright-field microscope images utilize defocusing to increase image contrast. The latter is related to the physical light phase through the transport of intensity equation (TIE). Recently, it was shown that it is possible to approximate the solution of the TIE using a low-pass monogenic signal framework. The purpose of this paper is to show that using the local phase of the aforementioned monogenic signal instead of the defocused image improves the cell/background classification accuracy.
The paper statement was tested on an image database composed of three cell lines: adherent CHO, adherent L929, and Sf21 in suspension. Local phase and local energy images were generated using the low-pass monogenic signal framework with axial derivative images as input. Machine learning was then employed to investigate the discriminative power of the local phase. Three classifier models were utilized: random forest (RF), support vector machine (SVM) with a linear kernel, and SVM with a radial basis function (RBF) kernel.
The improvement, averaged over cell lines, of classifying 5×5 sized patches extracted from the local phase image instead of the defocused image was 7.3% using the RF, 11.6% using the linear SVM, and 10.2% when a RBF kernel was employed instead of the linear one. Furthermore, the feature images can be sorted by increasing discriminative power as follows: at-focus signal, local energy, defocused signal, local phase. The only exception to this order was the superiority of local energy over defocused signal for suspended cells.
Local phase computed using the low-pass monogenic signal framework considerably outperforms the defocused image for the purpose of pixel-patch cell/background classification in bright-field microscopy.
几种处理明场显微镜图像的细胞检测方法利用离焦来提高图像对比度。后者与通过强度传输方程（TIE）的物理光相位有关。最近，已经表明可以使用低通单态信号框架来近似 TIE 的解。本文的目的是表明，使用上述单态信号的局部相位而不是离焦图像可以提高细胞/背景分类的准确性。
本文的陈述在由三种细胞系组成的图像数据库上进行了测试：贴壁 CHO、贴壁 L929 和悬浮 Sf21。使用轴向导数图像作为输入的低通单态信号框架生成局部相位和局部能量图像。然后使用机器学习来研究局部相位的判别能力。使用了三种分类器模型：随机森林（RF）、带有线性核的支持向量机（SVM）和带有径向基函数（RBF）核的 SVM。
平均而言，从局部相位图像而不是离焦图像提取的 5×5 大小的斑块进行分类时，RF 提高了 7.3%，线性 SVM 提高了 11.6%，而使用 RBF 核代替线性核时提高了 10.2%。此外，特征图像可以按增加的判别能力排序如下：聚焦信号、局部能量、离焦信号、局部相位。唯一的例外是局部能量对于悬浮细胞的离焦信号具有优越性。
使用低通单态信号框架计算的局部相位在明场显微镜中用于像素斑块细胞/背景分类的目的上，明显优于离焦图像。
