Sejuti Zarin Anjuman, Islam Md Saiful
Department of Electronics & Telecommunication Engineering, Chittagong University of Engineering & Technology, Chattogram, Bangladesh.
Sens Int. 2023;4:100229. doi: 10.1016/j.sintl.2023.100229. Epub 2023 Jan 31.
The novel coronavirus is the new member of the SARS family, which can cause mild to severe infection in the lungs and other vital organs like the heart, kidney and liver. For detecting COVID-19 from images, traditional ANN can be employed. This method begins by extracting the features and then feeding the features into a suitable classifier. The classification rate is not so high as feature extraction is dependent on the experimenters' expertise. To solve this drawback, a hybrid CNN-KNN-based model with 5-fold cross-validation is proposed to classify covid-19 or non-covid19 from CT scans of patients. At first, some pre-processing steps like contrast enhancement, median filtering, data augmentation, and image resizing are performed. Secondly, the entire dataset is divided into five equal sections or folds for training and testing. By doing 5-fold cross-validation, the generalization of the dataset is ensured and the overfitting of the network is prevented. The proposed CNN model consists of four convolutional layers, four max-pooling layers, and two fully connected layers combined with 23 layers. The CNN architecture is used as a feature extractor in this case. The features are taken from the CNN model's fourth convolutional layer and finally, the features are classified using K Nearest Neighbor rather than softmax for better accuracy. The proposed method is conducted over an augmented dataset of 4085 CT scan images. The average accuracy, precision, recall and F1 score of the proposed method after performing a 5-fold cross-validation is 98.26%, 99.42%,97.2% and 98.19%, respectively. The proposed method's accuracy is comparable with the existing works described further, where the state of the art and the custom CNN models were used. Hence, this proposed method can diagnose the COVID-19 patients with higher efficiency.
新型冠状病毒是严重急性呼吸综合征家族的新成员，可导致肺部以及心脏、肾脏和肝脏等其他重要器官出现轻度至重度感染。为了从图像中检测新型冠状病毒肺炎（COVID-19），可以采用传统的人工神经网络（ANN）。该方法首先提取特征，然后将这些特征输入到合适的分类器中。由于特征提取依赖于实验者的专业知识，因此分类率不是很高。为了解决这一缺点，提出了一种基于卷积神经网络（CNN）和K近邻（KNN）的混合模型，并采用5折交叉验证来对患者的CT扫描图像进行COVID-19或非COVID-19分类。首先，执行一些预处理步骤，如对比度增强、中值滤波、数据增强和图像缩放。其次，将整个数据集分成五个相等的部分或折用于训练和测试。通过进行5折交叉验证，可以确保数据集的泛化能力，并防止网络过拟合。所提出的CNN模型由四个卷积层、四个最大池化层和两个全连接层组成，共23层。在这种情况下，CNN架构用作特征提取器。从CNN模型的第四个卷积层提取特征，最后，使用K近邻而非softmax进行分类以获得更高的准确率。所提出的方法在一个包含4085张CT扫描图像的增强数据集上进行。在进行5折交叉验证后，所提出方法的平均准确率、精确率、召回率和F1分数分别为98.26%、99.42%、97.2%和98.19%。所提出方法的准确率与进一步描述的现有工作相当，现有工作使用了最先进的方法和自定义的CNN模型。因此，所提出的方法可以更高效地诊断COVID-19患者。
