基于ResNet101多特征融合的新型冠状病毒感染图像分类方法

Novel Method for Classifying Coronavirus Infection Images Based on ResNet101 Multi-feature Fusion

新型冠状病毒感染自爆发以来一直威胁着人类健康.针对现有对新型冠状病毒医学图像分类方法忽略了对深度神经网络前层特征的利用以及特征信息丢失不利于小区域的分类问题,本文提出了一种多特征融合利用的BAFPN-ResNet101模型.该模型以ResNet101为主干网络,引入了使用双线性插值法和通道注意力改进的特征金字塔结构,在主干网络与特征金字塔结构横向连接中使用通道注意力增强特征赋予不同的权重,然后利用特征金字塔结构将高层特征与低层特征融合.在公开数据集Chest X-ray(Covid-19&Pneumonia)上测试,实验结果显示,BAFPN-ResNet101模型在三分类实验中对识别新型冠状病毒感染胸部X射线的准确率、精确率、召回率分为97.41%、98.36%、97.20%.与其他方法相比,本文所提方法有效的利用了神经网络前层特征,对新型冠状病毒感染胸部X射线图像能够精确的识别,具有良好的泛化能力和性能.

Novel coronavirus pneumonia has been a threat to human health since its outbreak.To address the problem that existing methods for classifying medical images of novel coronaviruses neglect the use of deep neural network front layer features and the loss of feature information is detrimental to the classification of small regions,this paper proposes a BAFPN-ResNet101 model with multi-feature fusion utilization.The model uses ResNet101 as the backbone network,introduces a feature pyramid structure improved using bilinear interpolation and channel attention,uses channel attention to augment features with different weights in the lateral connection between the backbone network and the feature pyramid structure,and then uses the feature pyramid structure to fuse higher-level features with lower-level features.Tested on the publicly available dataset Chest X-ray(Covid-19&Pneumonia),the experimental results showed that the accuracy,precision and recall of the BAFPN-ResNet101 model for identifying novel coronavirus infected chest X-ray in the triple classification experiment were classified as 97.41%,98.36%and 97.20%.Compared with other methods,the proposed method in this paper effectively utilizes the neural network pre-layer features and can accurately identify the novel coronavirus-infected chest X-ray images with good generalization ability and performance.