Kaya Mustafa, Eris Mustafa
Firat University, Faculty of Technology, Digital Forensic Eng. Dept, Türkiye.
Biomed Signal Process Control. 2023 Apr;82:104559. doi: 10.1016/j.bspc.2022.104559. Epub 2023 Jan 2.
Covid-19 is one of the biggest global epidemics seen in the world in recent years. Because of this, people's daily lifestyles, the economic conditions of countries and individuals, and most importantly, their health status has been adversely affected all over the world. Millions of people around the world have died from this disease. For this reason, rapid and accurate detection of the disease is of great importance in terms of treatment and precautions. In addition, it is especially important to correctly distinguish between Covid-19 and non-Covid-19 pneumonia diseases for correct diagnosis and treatment. These two diseases cause similar symptoms, and the symptoms and the effects of the disease on the body should be carefully examined for their differentiation. Chest X-ray images, chest computerized tomography, and swab tests are commonly used to detect patients infected with COVID-19. This disease affects the lungs the most in the body and causes fatal side effects such as shortness of breath. Therefore, medical images taken from the chest play an important role in the diagnosis of the disease. The fact that X-rays are faster and cheaper than computerized tomography has led to an increase in studies on the detection of disease with X-rays. In recent years, the impressive results of deep learning in the field of computer vision have attracted researchers to this field when working with image data. This study aims to detect these diseases on chest X-ray images collected from Covid-19 patients, pneumonia patients, and healthy individuals. We proposed a hybrid feature extraction network namely DSENET which consists of DarkNet53, DarkNet19, DenseNet201, SqueezeNet, and EfficientNetb0. After a balanced data set was prepared, feature vectors were obtained from images using deep learning-based CNN models and the size of feature vectors was reduced by feature selection algorithms. Obtained features were classified by traditional machine learning methods such as SVMs. The number of features to be selected was tested by the iterative increment method and the parameters with the highest accuracy rate were obtained. As a result, it was seen that healthy and infected individuals were detected in 3 classes with an accuracy rate of 98.78%. In addition, the confusion matrix, precision, recall values, and F1 score of the obtained model are also given.
新冠病毒病是近年来全球出现的最大规模的流行病之一。正因如此，人们的日常生活方式、国家和个人的经济状况，以及最重要的，世界各地人们的健康状况都受到了不利影响。全球数百万人死于这种疾病。因此，快速准确地检测该疾病对于治疗和预防而言至关重要。此外，正确区分新冠病毒病和非新冠病毒病肺炎对于正确诊断和治疗尤为重要。这两种疾病会引发相似的症状，需要仔细检查症状以及疾病对身体的影响以进行区分。胸部X光图像、胸部计算机断层扫描和拭子检测是检测新冠病毒病感染者常用的方法。这种疾病对人体肺部影响最大，并会导致诸如呼吸急促等致命的副作用。因此，胸部医学影像在该疾病的诊断中发挥着重要作用。X光比计算机断层扫描更快且更便宜，这使得利用X光检测疾病的研究有所增加。近年来，深度学习在计算机视觉领域取得的显著成果吸引了研究人员在处理图像数据时投身该领域。本研究旨在从新冠病毒病患者、肺炎患者和健康个体收集的胸部X光图像上检测这些疾病。我们提出了一种混合特征提取网络，即DSENET，它由DarkNet53、DarkNet19、DenseNet201、SqueezeNet和EfficientNetb0组成。在准备好平衡数据集后，使用基于深度学习的卷积神经网络模型从图像中获取特征向量，并通过特征选择算法减小特征向量的大小。所获得的特征通过支持向量机等传统机器学习方法进行分类。通过迭代增量法测试要选择的特征数量，并获得准确率最高的参数。结果显示，在三类中检测健康个体和感染个体的准确率为98.78%。此外，还给出了所得模型的混淆矩阵、精确率、召回值和F1分数。
