Hurst R E, Bonner R B, Ashenayi K, Veltri R W, Hemstreet G P
Department of Urology, University of Oklahoma Health Sciences Center, Oklahoma City 73190, USA.
Cytometry. 1997 Jan 1;27(1):36-42.
We report on preliminary investigations of the use of an image analysis system to perform preliminary algorithmic classification of images of fluorochrome-labeled cells followed by capture of gray-level images of potentially abnormal cells for analysis by a neural network. Cells were labeled with an antibody against a bladder cancer tumor-associated antigen, and the neural net was used to distinguish true-positive cells from negative cells, false-positive cells (autofluorescent or nonspecific labeling), and cell-sized artifacts. Gray-level cell images were digitized and processed for analysis by a feed-forward neural network using back-propagation. The network was trained and tested with two independent image sets. Various network configurations and activation functions were investigated, including a sinusoidal activation function. At high power, the network agreed completely with the human observer's classification. At low power, a strong clustering of cells classified by the network with expert classification was seen, while the neural network showed roughly 75% concordance with the human observer. In addition, a set of four features extracted from raw cell images were investigated. The features were: shape factor, texture, area, and average pixel intensity. A network trained with these features performed better than one operating with gray-level images. We conclude that using neural networks to recognize and classify images captured by an image analysis microscope is feasible.
我们报告了一项初步研究，该研究利用图像分析系统对荧光染料标记细胞的图像进行初步算法分类，随后捕获潜在异常细胞的灰度图像，以供神经网络分析。细胞用针对膀胱癌肿瘤相关抗原的抗体进行标记，神经网络用于区分真阳性细胞与阴性细胞、假阳性细胞（自发荧光或非特异性标记）以及细胞大小的伪像。灰度细胞图像被数字化，并使用反向传播算法通过前馈神经网络进行处理以进行分析。该网络使用两个独立的图像集进行训练和测试。研究了各种网络配置和激活函数，包括正弦激活函数。在高倍镜下，该网络与人类观察者的分类完全一致。在低倍镜下，可以看到经网络分类的细胞与专家分类呈现出很强的聚类现象，而神经网络与人类观察者的一致性约为75%。此外，还研究了从原始细胞图像中提取的一组四个特征。这些特征是：形状因子、纹理、面积和平均像素强度。用这些特征训练的网络比使用灰度图像的网络表现更好。我们得出结论，使用神经网络识别和分类由图像分析显微镜捕获的图像是可行的。
