The overlapping imaging characteristics of COVID-19 viral pneumonia and non-COVID-19 viral pneumonia chest X-rays (CXRs) make differentiation difficult for radiologists. Machine learning (ML) has demonstrated promising outcomes in a range of medical sectors, enhancing diagnostic accuracy through its interaction with radiological tests. The potential contribution of ML models in assisting radiologists in discriminating COVID-19 from non-COVID-19 viral pneumonia from CXRs, on the other hand, deserves further examination and exploration. The goal of this study is to empirically assess ML models' capacity to classify X-ray images into COVID-19, pneumonia, and normal cases. The study evaluates the efficacy of K-nearest Neighbor (KNN), random forest (RF), AdaBoost (AB), and neural networks (NN) with various hidden neuron configurations using a wide range of performance measures. These metrics evaluate the area under the curve (AUC), classification accuracy (CA), F1 score (F1), precision, and recall, resulting in a comprehensive evaluation technique. ROC analysis is used to gain a thorough knowledge of the models' discriminating skills. The results show that NN models, particularly those with 100 and 150 hidden neurons, outperform in all criteria, proving their ability to reliably categorize medical disorders. Notably, the study emphasizes the difficulties in separating COVID-19 from pneumonia, emphasizing the importance of strong classification methods. While the study provides useful insights, its drawbacks include the use of a single dataset, the absence of more sophisticated deep learning architectures, and a lack of interpretability analyses. Nonetheless, the study adds to the developing area of medical picture categorization, directing future attempts to improve diagnosis accuracy and widen the use of machine learning in healthcare. The findings highlight the utility of NN models in medical diagnostics and pave the way for future study in this vital area of technology and healthcare.
Top1. Introduction
The infectious disease termed COVID-19 is caused by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). This virus induces pronounced respiratory symptoms and was initially identified in Wuhan, China, in 2019. Following its rapid global dissemination, it was officially designated as a worldwide pandemic (Cucinotta and Vanelli, 2020). As of October 15, 2021, the COVID-19 Dashboard maintained by John Hopkins University has documented an estimated 240 million confirmed cases of infection and approximately 4.8 million recorded fatalities on a global scale (Lauer et al., 2020; Jaiswal et al., 2021). The cardinal manifestations of COVID-19 encompass elevated body temperature, fatigue, respiratory distress, and impairment of gustatory or olfactory senses. In instances of heightened severity, the disease can give rise to respiratory complications, particularly pneumonia, warranting hospitalization and occasionally necessitating admission into intensive care units (Lauer et al., 2020).
Distinguishing between COVID-19 pneumonia and typical pneumonia based on clinical attributes can present challenges. Both conditions exhibit akin signs and symptoms, including fever, fatigue, non-productive cough, and respiratory distress. The elevated morbidity and mortality rates linked to COVID-19 pneumonia have imposed a substantial burden on healthcare infrastructures (Jaiswal et al., 2021). To curb the propagation of the pandemic and optimize the allocation of medical resources, it is imperative to swiftly diagnose and isolate individuals afflicted with either common pneumonia or COVID-19.
Despite the convergence of symptoms and diagnostic complexities, the radiographic representations obtained through Computed Tomography (CT) for general pneumonia and COVID-19 manifest resemblances. This overlapping imagery further compounds the intricacy of effectively discerning between the two maladies (Cheng et al., 2019).
The identification of COVID-19 is commonly accomplished through real-time polymerase chain reaction (RT-PCR) testing for the presence of the SARS-CoV-2 virus. While RT-PCR exhibits commendable specificity for COVID-19, its sensitivity for accurately detecting cases of the disease has been relatively lower (Nishio et al., 2020). Notably, chest computed tomography (CT) scans have demonstrated utility in identifying atypical conditions associated with COVID-19 pneumonia. The distinctive CT findings pertinent to COVID-19 offer a basis for differentiation from viral and bacterial pneumonia presentations (Bai et al., 2020).
Numerous recent investigations have aimed to differentiate between COVID-19 and pneumonia, each harboring inherent limitations and strengths. This study, however, presents an empirical scrutiny of diverse machine learning (ML) models, including K-nearest Neighbor (KNN), Random Forest (RF), AdaBoost (AB), and Neural Networks (NN) with concealed layers containing 100 and 150 neurons. Employing a dataset sourced from the Kaggle repository, these models are evaluated utilizing state-of-the-art performance metrics encompassing the area under the curve (AUC), classification accuracy (CA), F1 Score (F1), precision, and recall.
The study makes a substantial contribution to the domain of medical image classification by addressing the crucial challenge of classifying individuals into COVID-19, Pneumonia, and Normal categories based on X-ray images. The research undertakes a comprehensive empirical analysis to evaluate the effectiveness of several machine learning (ML) models, encompassing K-nearest Neighbor (KNN), Random Forest (RF), AdaBoost (AB), and Neural Network (NN) with various configurations. This meticulous investigation is underpinned by a range of well-established evaluation metrics, including Area under the Curve (AUC), Classification Accuracy (CA), F1 Score, accuracy, and recall. These metrics collectively provide a comprehensive understanding of the strengths and limitations intrinsic to each model's diagnostic capacities for medical conditions.