融合宫颈细胞领域特征的多流卷积神经网络分类算法

Cervical Cell Features Based Multi-Stream Convolutional Neural Networks Classification Method

细胞分类是宫颈癌计算机辅助诊断研究和应用的关键技术.针对通用深度学习分类算法在细胞分类中缺少领域知识指导这一局限性,提出一种基于数据驱动和宫颈细胞领域知识的多流卷积神经网络分类算法.文中算法以细胞和细胞核图像为输入,通过卷积神经网络提取图像特征,并根据宫颈细胞标准分级系统中领域知识提取人工设计特征,最后将上述2种特征进行拼接,并经过全连接层融合,构建适用于细胞分类的多流卷积神经网络.实验结果表明,文中算法在仅使用Alexnet作为基础网络的情况下,在Herlev宫颈细胞图像数据集上的正常与异常细胞的分类准确率达到99%,取得了该数据库上目前最好的分类结果;在Ideepwise数据集上,按照细胞学诊断报告的分级准确率为85%,相比单流网络提升3%.

Cervical cell classification is a key problem for computer-aided cervical screening applications. Previous methods mainly focus on designing data-driven methods to learn classification models from labeled dataset, and the domain knowledge of cervical cells problem is seldom explored to further improve the classification performance. To address this problem, we propose a multi-stream convolutional neural network classification algorithm built upon data-driven approach as well as cervical cell domain knowledge. The algorithm uses cell image, nucleus image and artificial designed features based on domain knowledge in the bethesda system(TBS) as input, and extracts multi-stream features through convolutional neural network.Finally, the above three-stream features are fused and then fed to a classification model to give a classification prediction. Experiments show that the classification accuracy of proposed algorithm based on Alexnet is99% on squamous epithelial cells from Herlev cervical cell database, the best reported classification performance on the database up to now. The method proposed achieves the classification accuracy of 85% on our Ideepwise database based on the TBS2014 standard, which gets a gain of 3% compared to the baseline single-stream networks.