注意力特征融合VGG-16网络的胸透病理识别方法

Attention Feature Fusion with VGG-16 Network for Chest X-Ray Pathologic Identification

美国国立卫生研究院发布了一个名为“Chest-ray8”的胸部X光图像数据库,这些图像共标记八种类型的疾病。从临床图像中识别病理,开发人工智能辅助诊断系统,帮助医生识别病理一直是一项具有挑战性的任务,为更好地解决这一问题,本文提出注意力特征融合VGG-16网络的胸透病理识别方法。首先,X光图像通过VGG-16特征提取网络得到特征向量;然后,特征向量分别通过通道注意力机制得到不同通道加权系数的特征向量A、空间注意力机制得到不同区域系数的特征向量B,将特征向量A和特征向量B以及第一层卷积层输出的特征向量C进行向量拼接;最后将拼接的向量输入全连接层输出分类结果。实验结果证明,本方法实现多分类正确率达到75.94%,相比基线VGG-16高3.75%;和其他使用“Chest-ray8”数据库进行二分类病理识别方法相比,本文模型实现了病理的多分类问题。为深度学习技术应用于胸部X光不同病理的计算机辅助诊断系统提供了可行性。

The National Institutes of Health publishes a database of Chest X-ray images called Chest-ray8, which are labeled for a total of eight types of disease. It has always been a challenging task to iden-tify pathology from clinical images and develop an AI-assisted diagnosis system to help doctors identify pathology. In order to better solve this problem, this paper proposed a chest X-ray patho-logic recognition method based on the fusion of attention features and the VGG-16 network. Firstly, the X-ray image feature vectors were obtained by the VGG-16 feature extraction network. Then, feature vectors A with different channel weighting coefficients were obtained by the channel atten-tion mechanism, and feature vectors B with different regional coefficients were obtained by the spa-tial attention mechanism respectively. Feature vectors A, B and C output from the first convolution layer were spliced. Finally, the spliced vector is input into the full connection layer to output the classification result. Experimental results show that the accuracy of multi-classification achieved by this method is 75.94%, which is 3.75% higher than that of baseline VGG-16. Compared with other methods using Chest-ray8 database for binary pathological identification, the model in this paper achieves the problem of multi-classification of pathology. It provides the feasibility of applying deep learning technology to computer-aided diagnosis systems of different pathologies of chest X-rays.