基于生成对抗网络的肺结节良恶性诊断算法

针对实性肺结节CT影像数据量少、人工标注耗时耗力等问题，提出一种结合生成对抗网络和集成学习的实性肺结节良恶性计算机辅助诊断方法.首先，使用基于梯度惩罚的生成对抗网络对肺结节CT影像数据集进行扩充，缓解由数据量少、样本类别不均衡导致的模型过拟合.然后，利用卷积神经网络进行CT影像特征提取，并通过主成分分析对深度特征进行降维.最后，联合CT图像特征和有效临床信息，采用集成学习方法构建分类模型预测实性肺结节良恶性.基于多中心临床数据分析表明，相比于传统卷积神经网络模型，所提出方法有更好的预测性能.

Diagnosis Algorithm of Pulmonary Nodules Malignancy Based on Generative Adversarial Network

To solve the problem of data scarcity and expensive costs in manual labeling of CT images of solid pulmonary nodules， a computer aided diagnosis algorithm for classification between lung tuberculosis and lung adenocarcinoma of solid pulmonary nodules by combining generative adversarial network and ensemble learning was proposed. Firstly， the original CT image dataset was augmented using Wasserstein generative adversarial network with gradient penalty(WGAN-GP)， in order to relieve the problem of overfitting caused by small-scale dataset and class imbalance. Then， feature extraction was performed using convolutional neural network， followed by a dimension reduction procedure using principal component analysis(PCA). Finally， deep features concatenated with effective subjective features were classified by an ensemble learning model to give final prediction of the patient.Analysis based on multi-center clinical data indicated that the proposed algorithm has better performance compared to traditional convolutional neural network method.