基于深度学习的腹膜腔脱落细胞病理图像识别

Pathology image recognition of peritoneal exfoliative cells based on deep learning

针对体液细胞病理诊断自动分类识别的问题,构建一种基于深度卷积神经网络的自动化识别模型(CNN-LS)。首先对图像样本进行灰度级转换、ZCA白化、归一化与标注处理,降低图像特征间的相关性与数据冗余。其次,在CNN-LS模型构建过程中引入改进的激活函数(LReLU-Softplus)用于提高模型的收敛速度和避免可能出现的饱和非线性问题,并通过实验验证获取CNN-LS模型的最佳卷积核数量和尺寸大小。最后将CNN-LS与CS+SVM,PCA+QSOFM,ANN,CNN这4种分类方法做性能对比。实验表明CNN-LS模型在针对腹膜腔脱落细胞病理图像的癌细胞分类识别过程中具有较明显的优势。

To solve the problem of automatic classification and recognition of pathological diagnosis of humoral cells,in this study,a kind of automatic recognition model was built based on the depth of convolution neural network.First,the image samples were processed with a grayscale conversion,ZCA whitening,normalization and labeling to reduce the correlation between image features and data redundancy.Then an improved activation function(LReLU-Softplus)was introduced in the construction of CNN-LS model to improve the speed of convergence and avoid possible saturated nonlinear problems,and obtained the optimal number and size of convolution kernel of CNN-LS model through experimental verification.Finally,performance contrast were made among the four classification methods of CS+SVM,PCA+QSOFM,ANN and CNN.Experimental results showed that CNN-LS model has obvious advantages in the process of classification and identification of cancer cells in the pathologic images of peritoneal cavity shedding cells.