COVID-19 classification from ground glass opacities of radiologic images via pytorch convolutional neural network

• Main Authors: Que, Carlos Joshua V., Vjuan, Macario S., Apostol, Alyanna Teresa Q.
• Format: text
• Language: English
• Published: Animo Repository 2022
• Subjects: Image processing; Programming languages (Electronic computers); COVID-19 (Disease); Physics
• Online Access: https://animorepository.dlsu.edu.ph/etdb_physics/16; https://animorepository.dlsu.edu.ph/context/etdb_physics/article/1018/viewcontent/2022_Apostol_Que_Vjuan_COVID_19_Classification_from_Ground_Glass_Opacities_of_Radiologic_Full_text.pdf
• Institution: De La Salle University

The researchers performed COVID-19 classification from the ground glass opacities (GGOs) of radiologic images of suspected patients by convolutional neural network (CNN) program using Pytorch Tensor, a python script programming language data type. Due to the increasing number of COVID -19 variants, efficient, accurate, and early diagnosis is important to provide proper treatment plans and management of patients. Aside from qRT-PCR (quantitative reverse transcription - polymerase chain reaction) diagnostic tests, analysis of chest radiographic image helps in diagnosing patients who developed pneumonia due to COVID-19. In recent years, artificial intelligence frameworks, particularly CNNs have achieved remarkable growth in medical image analysis and case classification for a variety of medical conditions. In this study, the researchers applied the principle of neural network supervised learning in classifying COVID-19 patients based on their characteristic ground glass opacity radiologic images. In order to implement this algorithm, a convolutional neural network program is developed using a python script programming language to characterize and classify positive COVID-19 chest x-ray scans by analyzing GGOs radiologic images using a Pytorch Tensor data type. The CNN model consisted of 300 x-ray scans in which 150 were normal as diagnosed by a physician and the other half as diagnosed by the physician with COVID-19. All of the images were normalized to a 299 x 299 pixel array size which were then divided into two groups consisting of 200 images for the training data set of CNN and 100 images for test classification. GGO radiologic images were used in three batch intervals and a subsequent number of x-ray scans were added. Twenty five (25) x-ray scan images were added to the testing dataset while 50 scans were added to the training dataset for every batch interval. Experimental results showed that the CNN model was able to classify x-ray image scans of people who are diagnosed as normal and people who are positive with COVID-19 with an accuracy of 89.20%, sensitivity of 83.33%, specificity of 97.17%, precision of 97.56%, and F-measure of 89.88%. Receiver Operating Characteristic Curve Analysis showed that the CNN model for both training and testing batches are all above the random classifier line indicating that the model created may be able to classify GGO X-ray images correctly. For training and testing batches, the true positive rate is 0.957 and 0.85, whereas the false positive rate is 0.115 and 0.06 respectively. To contextualize the CNN model, it would be ideal to use GGO radiologic images coming from the Philippines rather than data sets coming from Qatar and Bangladesh in training the CNN model.