基于SPD多尺度输入的ST-MASA的肺炎智能检测模型  

Intelligent Pneumonia Detection Model of ST-MASA Based on SPD Multi-scale-input

作　　者：李芳芳 束建华 阚峻岭 殷云霞 孙大勇 马春 LI Fangfang;SHU Jianhua;KAN Junling;YIN Yunxia;SUN Dayong;MA Chun(School of Medical Information Engineering,Anhui University of Traditional Chinese Medicine,Hefei 230012,China)

机构地区：[1]安徽中医药大学医药信息工程学院,安徽合肥230012

出　　处：《宿州学院学报》2023年第12期11-17,共7页Journal of Suzhou University

基　　金：安徽省高等学校自然科学重点研究项目(2022AH050475,KJ2020A0394);安徽中医药大学自然科学重点研究项目(2020zrzd20,2020zrzd19,2020zrzd17,2021zrzd12);安徽省省级教学研究项目(2020jyxm1018,2020jyxm1020)。

摘　　要：肺部X光片的临床诊断结果可以作为新冠肺炎及其他肺炎诊断的重要依据,而X光片所显示的肺炎病变的相似性及阅片量巨大,医生传统的阅片识别存在误诊、漏诊和时间消耗等问题。因此,提出了一种融合空间金字塔分解(Spatial pyramid decomposition, SPD)模块进行多尺度输入的ST-MASA(Swin Transformer with Multi-Head Axial-Self-Attention)的肺炎智能检测模型,用于COVID-19和多类型肺炎的自动分类。该模型能够自动关注肺炎病灶的判别信息和多尺度特征,进而更好地进行COVID-19、肺不透明(Lung_Opacity)、非COVID的病毒性肺炎(Viral_Pneumonia)和正常(Normal)的X光片进行分类,以便更好地帮助放射科医生进行医疗诊断工作。实验结果表明,所提出的模型在准确率、召回率、F1-Measure等指标上均优于经典的网络模型ResNet 50、ResNet 101、Inception net-V3和Swin Transformer。The clinical diagnosis results based on lung X-rays provide important evidence in the COVID-19 pneumonia diagnosis process and for some other pneumonia.However,due to the similarity of the lesions among many types of pneumonia displayed by X-rays,and due to the huge amount of X-ray film reading of a doctor′s daily work,there are some problems such as misdiagnosis,missed diagnosis and time consumption in traditional X-ray film rea-ding and identification.Therefore,an intelligent pneumonia detection model of ST-MASA(Swin Transformer with Multi-Head Axial-Self-Attention)that integrates the Spatial pyramid decomposition(SPD)module for multi-scale input is proposed for the automatic classification of COVID-19 and multi-type pneumonia.The model will also be able to automatically focus on the discrimination information and multi-scale characteristics of pneumonia lesions,and further better classify COVID-19,Lung_Opacity,non-covid viral pneumonia and Normal X-ray films so as to better help radiologists to carry out medical diagnosis work.The experimental results show that the proposed model is superior to the classical network models ResNet 50,ResNet 101,Inception net-V3 and Swin Transformer in terms of accuracy,recall rate and F1-Measure.

关 键 词：肺炎智能检测 空间金字塔分解 多尺度输入 多头轴向自注意力机制 Swin Transformer 

分 类 号：TP391.44[自动化与计算机技术—计算机应用技术]