Machine Learning-Based Computational Magnetic Resonance Models for Rapid Diagnosis of Corona Virus Disease 2019 (COVID-19) Using Magnetic Resonance Fingerprinting Data

Abstract

: Coronavirus 2019 (COVID-19) caused by Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) is a respiratory track disease characterized by fever, dry cough, fatigue, and gastrointestinal symptoms. Currently, there is no definite treatment for COVID-19 although some drugs are under investigation. To promptly identify patients and prevent further spreading, they must be diagnosed in time and extra care has to be taken to ensure healthcare providers are not infected at all. Early detection and diagnosis of COVID-19 increases the chances of recovery of patients and this can be achieved through the use of computed tomography (CT) and magnetic resonance imaging (MRI) scanning. Although chest CT and digital radiography (DR) is important in the diagnosis and treatment of COVID-19, MRI is crucial for not just diagnosis of the disease but also for detection of other complications. MRI examinations for COVID-19 patients, protection of medical personnel and disinfection of MRI equipment is a challenging due high contingency of the disease. In addition to these problems, MRI examinations are costly. In order to overcome these challenges, this study proposes a computational method based of magnetic resonance fingerprinting (MRF) data, Bloch NMR response and machine learning. Clinical MRF were obtained from a recent experimental study on COVID-19. Since the specific MRF measurements in individual patients are not currently available, R computer codes were developed to simulate the measurements per subject. Taking the simulated data as individual data points, the time-independent NMR Bloch flow equation was solved analytically (as shown in equations (1) to (6)) and employed to compute the MRI signal for each subject based on spleen analysis data. The simulations were also done for cardiac MRI, hepatic panel test and brain volume data. The dataset generated from these simulations were then splitted into two (80% as training set and 20% as test set). As shown in figure 1, machine learning algorithms were then used to train the dataset. Support vector machine, logistic regression, decision tree and gradient boosting classifier returned accuracies of 88%, 82%, 88% and 88% respectively for spleen analysis dataset. For the hepatic panel test, cardiac MRI and brain volume datasets, these algorithms all returned 100% accuracies. Logistic regression model was then selected to do deploy the machine learning algorithm with streamlit and Heroku.