Multiclass Convolution Neural Network for Classification of COVID-19 CT Images. (28th April 2022)

• Record Type: Journal Article
• Authors: Woan Ching, Serena Low; Lai, Khin Wee; Chuah, Joon Huang; Hasikin, Khairunnisa; Khalil, Azira; Qian, Pengjiang; Xia, Kaijian; Jiang, Yizhang; Zhang, Yuanpeng; Dhanalakshmi, Samiappan
• Other Names: Ding Bai Yuan Academic Editor.
• Is Part Of: Computational intelligence and neuroscience. Volume 2022(2022)
• Publication Date: 2022-04-28
• Subjects: Neurosciences -- Data processing -- Periodicals; Computational intelligence -- Periodicals; Computational neuroscience -- Periodicals; 612.80285
• DOI: 10.1155/2022/9167707
• Languages: English
• ISSNs: 1687-5265

Abstract

Abstract : In the late December of 2019, a novel coronavirus was discovered in Wuhan, China. In March 2020, WHO announced this epidemic had become a global pandemic and that the novel coronavirus may be mild to most people. However, some people may experience a severe illness that results in hospitalization or maybe death. COVID-19 classification remains challenging due to the ambiguity and similarity with other known respiratory diseases such as SARS, MERS, and other viral pneumonia. The typical symptoms of COVID-19 are fever, cough, chills, shortness of breath, loss of smell and taste, headache, sore throat, chest pains, confusion, and diarrhoea. This research paper suggests the concept of transfer learning using the deterministic algorithm in all binary classification models and evaluates the performance of various CNN architectures. The datasets of 746 CT images of COVID-19 and non-COVID-19 were divided for training, validation, and testing. Various augmentation techniques were applied to increase the number of datasets except for testing images. The images were then pretrained using CNN to obtain a binary class. ResNeXt101 and ResNet152 have the best F1 score of 0.978 and 0.938, whereas GoogleNet has an F1 score of 0.762. ResNeXt101 and ResNet152 have an accuracy of 97.81% and 93.80%. ResNeXt101, DenseNet201, and ResNet152 have 95.71%, 93.81%, and 90% sensitivity, whereas ResNeXt101, ResNet101, and ResNet152 have 100%, 99.58%, and 98.33 specificity, respectively.