基于M超图像的气胸自动诊断方法  

Automatic Diagnosis Method of Pneumothorax Based on M-Ultrasound Image

作　　者：马一博[1] 陈益 刘思言 闫士举[2] 

机构地区：[1]苏州大学附属第三医院超声医学科,江苏 常州 [2]上海理工大学健康科学与工程学院,上海

出　　处：《建模与仿真》2023年第4期3512-3521,共10页Modeling and Simulation

摘　　要：目的:气胸是一种可危及生命的呼吸急症,超声是气胸诊断常用方法。本文研究基于机器学习的M模式超声图像(M超)分类,以辅助医生进行气胸诊断。方法:采用包括肺滑、肺点和肺滑消失三类样本在内共600幅M超图像,由超声医学科医生划分为典型和非典型两个子集。提取出图像的灰度特征、LBP特征、GLCM特征和HOG特征,对特征数据降维,再运用SVM、逻辑回归、XGBoost、LGBM、随机森林五种算法进行图像分类。结果和结论:在典型数据集上,灰度特征 + 逻辑回归分类算法组合的分类效果最佳,其准确度、特异性、敏感性分别为99%、0.9714、0.99;在非典型数据集上,灰度特征 + SVM分类算法组合的分类效果最佳,其准确度、特异性、敏感性分别为98.33%、0.9714、0.99;在混合数据集上,灰度特征 + SVM分类算法组合的分类效果最佳,其准确度、特异性、敏感性分别为94.58%、0.8417、0.95,故采用灰度特征 + SVM分类算法组合对M超图像进行分析有助于辅助医生进行气胸诊断。创新点:通过尝试多种特征提取算法和分类算法的不同组合找出了适合进行气胸自动诊断的算法组合。Objective: Pneumothorax is a life-threatening respiratory emergency. Ultrasound is a common di-agnostic method for pneumothorax. This paper studies M-mode ultrasound image classification based on machine learning to assist doctors in pneumothorax diagnosis. Data and METHODS: A total of 600 M ultrasound images were used, including three types of samples of lung slip, lung spot and disappearance of lung slip, which were divided into two subsets, typical and atypical, by doctors in the department of ultrasound medicine. The gray feature, LBP feature, GLCM feature and HOG fea-ture of the image are extracted, the dimension of the feature data is reduced, and then SVM, logistic regression, XGBoost, LGBM and random forest algorithms are used for image classification. Results: On the typical data set, the combination of gray feature and logistic regression classification algo-rithm had the best classification effect, and its accuracy, specificity and sensitivity were 99%, 0.9714, 0.99, respectively. On the atypical data set, the combination of gray feature and SVM classi-fication algorithm has the best classification effect, and its accuracy, specificity and sensitivity are 98.33%, 0.9714 and 0.99, respectively. On the mixed data set, the combination of gray feature and SVM classification algorithm has the best classification effect, and its accuracy, specificity and sensi-tivity are 94.58%, 0.8417 and 0.95, respectively. Conclusion: The combination of gray feature and SVM classification algorithm to analyze M-ultrasound images is helpful to assist doctors in the diag-nosis of pneumothorax.

关 键 词：灰度特征 随机森林 机器学习 数据降维 图像分类 算法组合 超声图像 自动诊断 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]