摘要
针对胸部X射线图像的病变区域辨识度低、准确捕捉病变空间位置难等问题,提出了一种有利于提高胸片图像分类精度的多尺度注意力信息复用网络。首先,通过引入多路空间信息复用模块,增强疾病部位在特征图及通道之间的位置联系;其次,通过多尺度融合注意力模块,整合多尺度图像特征信息,自动捕捉病灶位置变化,以实现对关键病理信息的灵活关注;最后,通过非对称移位焦点损失函数,缓解胸部疾病样本分布不平衡的问题。在公开数据集ChestX-ray14和CheXpert上的多组实验表明:本文网络在两个数据集上的平均AUC值分别达到0.847和0.901,优于近年来较为先进的网络模型,表明该网络能有效地提高胸部疾病的分类精度。
To address issues such as low recognition of lesion areas in chest X-ray images and the difficulty in accurately capturing the spatial positions of lesions,a multi-scale attention information multiplexing network that helps improve the dassification accuracy of chest X-ray images was proposed in this paper.Firstly,by introducing multiple spatial information multiplexing blocks,the network enhances the positional connections between disease regions on feature maps and across channels;Secondly,through a multi-scale integration attention blocks,the network integrates multi-scale image feature information to automatically capture disease location variations and flexibly focus on key pathological information;Finally,the problem of imbalanced distribution of chest disease samples was alleviated by using an asymmetric shift focus loss function.Multiple experiments on the publicly available datasets ChestX-ray14 and CheXpert have shown that the average area under curve(AUC)value of the proposed network on two datasets reached 0.847 and 0.901 respectively,which is superior the more advanced network models in recent years.This indicates that the network can effectively improve the classification accuracy of chest diseases.
作者
张瑞峰
郭芳兆
李锵
ZHANG Rui-feng;GUO Fang-zhao;LI Qiang(School of Microelectronics,Tianjin University,Tianjin 300072,China)
出处
《吉林大学学报(工学版)》
北大核心
2025年第11期3686-3696,共11页
Journal of Jilin University:Engineering and Technology Edition
基金
国家自然科学基金项目(62071323)
超声医学工程国家重点实验室开放课题项目(2022KFKT004)
天津市自然科学基金重点项目(22JCZDJC00220).
关键词
计算机应用技术
胸部X光图像分类
空间信息复用
多尺度注意力
非对称移位焦点损失
computer application technology
chest X-ray image classification
spatial information multiplexing
multi-scale attention
asymmetric shift focus loss function
