用于肺炎图像分类的优化卷积神经网络方法

目前，卷积神经网络（CNN）开始应用在肺炎分类领域。针对层数较浅、结构较为简单的卷积网络对肺炎识别的准确率难以提高的情况，采用深度学习方法，并针对采用深度学习方法时常常需要消耗大量的系统资源，导致卷积网络难以在用户端部署的问题，提出一种使用优化的卷积神经网络的分类方法。首先，根据肺炎图像的特征，选择具有良好图像分类性能的AlexNet与InceptionV3模型；然后，利用医学影像特点对层次更深、结构更加复杂的InceptionV3模型进行预训练；最后，通过知识蒸馏的方法，将训练好的"知识"（有效信息）提取到AlexNet模型中，从而实现在减少系统资源占用的同时，提高准确率的效果。实验数据表明，使用知识蒸馏后，AlexNet模型的准确率、特异性与灵敏度分别提高了4.1、7.45、1.97个百分点，且对图像处理器（GPU）占用相比InceptionV3模型减小了51个百分点。

Optimized convolutional neural network method for classification of pneumonia images

Currently, Convolutional Neural Network (CNN) is applied in the field of pneumonia classification. Aiming at the hardness to improve the accuracy of pneumonia recognition of convolution network with shallow layers and simple structure, deep learning method was adopted; and concerning the problem that the deep learning method often consumes a lot of system resources, which makes the convolution network difficult to be deployed at user end, an classification method based on optimized convolution neural network was proposed. Firstly, according to the features of pneumonia images, AlexNet and Inception V3 models with good image classification performance were selected. Then, the characteristics of medical images were used to re-train the Inception V3 model with deeper layers and more complex structure. Finally, through knowledge distillation method, the trained "knowledge" (effective information) was extracted into AlexNet model, so as to reduce the occupancy of system resources and improve the accuracy. The experimental data show that after knowledge distillation, AlexNet model has the accuracy, specificity and sensitivity improved by 4.1, 7.45 and 1.97 percentage points respectively, and has the Graphics Processing Unit (GPU) occupation reduced by 51 percentage points compared with InceptionV3 model.