Application of Machine Learning in Diagnosis of COVID-19 Through X-Ray and CT Images: A Scoping Review

Coronavirus disease, first detected in late 2019 (COVID-19), has spread fast throughout the world, leading to high mortality. This condition can be diagnosed using RT-PCR technique on nasopharyngeal and throat swabs with sensitivity values ranging from 30 to 70%. However, chest CT scans and X-ray images have been reported to have sensitivity values of 98 and 69%, respectively. The application of machine learning methods on CT and X-ray images has facilitated the accurate diagnosis of COVID-19. In this study, we reviewed studies which used machine and deep learning methods on chest X-ray images and CT scans for COVID-19 diagnosis and compared their performance. The accuracy of these methods ranged from 76% to more than 99%, indicating the applicability of machine and deep learning methods in the clinical diagnosis of COVID-19.


INTRODUCTION
First identified in Wuhan, China, severe pneumonia caused by Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) quickly spread all over the world. The resultant disorder was named coronavirus disease (COVID-19) (1,2). COVID-19 has various clinical symptoms, including fever, cough, dyspnea, fatigue, myalgia, headache, and gastrointestinal complications (3)(4)(5). Diagnosis of COVID-19 infection through RT-PCR on nasopharyngeal and throat swab samples has been reported to yield positive results in 30-70% of cases (6,7). On the other hand, chest CT scans and X-ray images have been reported to have sensitivity values of 98 and 69%, respectively (7)(8)(9). The most typical radiological signs in these patients include multifocal and bilateral ground-glass opacities and consolidations, particularly in the peripheral and basal sites (10). However, interpretation of the results of these imaging techniques by expert radiologists might encounter some problems leading to reduced sensitivity (11). Artificial intelligence has recently gained the attention of both clinicians and researchers for the appropriate management of the COVID-19 pandemic (12). As an accurate method, artificial intelligence is able to identify abnormal patterns of CT and X-ray images. Using this method, it is possible to assess certain segment regions and take precise structures in chest CT images facilitating diagnostic purposes. Artificial intelligence methods have been shown to detect COVID-19 and distinguish this condition from other pulmonary disorders and community-acquired pneumonia (13). Both deep learning and machine learning approaches have been used to predict different aspects of the COVID-19 outbreak. Support vector and random forest are among the most applied machine learning methods, while Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Generative Adversarial Networks (GAN), and Residual Neural network are among the deep learning methods used in this regard (14). In this study, we reviewed studies which used machine and deep learning methods on chest X-ray images and CT scans for the purpose of COVID-19 diagnosis and compared their performance.

Search Strategy
The research question was: "What are the applications of machine learning techniques and their performances in COVID-19 diagnosis using X-ray images?". The search of the present review was based on the PICO elements, which were as follows: • P (Problem/Patient/Population): Patients' CT scans and Chest X-rays. In other words, we were looking for publications that evaluated the performance of any machine learning or deep learning approaches based on inclusion and exclusion criteria. Studies that used other types of medical image modalities (e.g., ultrasound images) were excluded. An electronic search was conducted on PubMed, Google Scholar, Scopus, Embase, arXiv, and medRxiv for finding the relevant literature. Duplicate studies were removed. Studies that were cited within the retrieved papers were reviewed for finding missing studies. For identifying proper journal papers and conference proceedings, investigators screened the title and abstracts based on inclusion and exclusion criteria independently. Finally, considering the inclusion and exclusion criteria, investigators identified the eligible publications in this stage independently.

Inclusion Criteria
The following inclusion criteria were used in the selection of the articles: (1) Studies that applied machine learning or deep learning algorithms, (2) Studies that evaluated the measurement of model outcomes in comparison with ground truth or gold standards, and (3) Studies that used algorithms to analyze radiographic images (CT scan, Chest X-ray, etc.).

Exclusion Criteria
The following studies were excluded: (1) Studies that used any machine learning or deep learning approaches for problems not directly related to the COVID-19 imaging, (2) Studies that used other artificial intelligence techniques or classic computer vision approaches, (3) Studies that did not provide a clear explanation of the machine learning or deep learning model that was used to solve their problem, and (4) Review studies. The latter were excluded as we did not aim to review the data on an original level without any second-hand interpretations (summation, inferences, etc.). Figure 1 shows the flowchart of the study design.

RESULTS
We obtained 105 studies that used machine or deep learning methods to assess chest images of COVID-19 patients. These studies have used different analytical methods. For instance, Ardakani et al. (15) have assessed radiological features of CT images obtained from patients with COVID-19 and non-COVID-19 pneumonia. They used decision tree, K-nearest neighbor, naïve Bayes, support vector machine, and ensemble classifiers to find the computer-aided diagnosis system with the best performance in distinguishing COVID-19 patients from non-COVID-19 pneumonia. They reported that site and distribution of pulmonary involvement, the quantity of the pulmonary lesions, ground-glass opacity, and crazy-paving as the most important characteristics for differentiation of these two sets of patients.  Tables 1, 2 summarize the features of studies which adopted machine learning methods in CT images and chest X-ray of COVID-19 patients, respectively.         Notably, the application of these methods on X-rays has offered promising results. Such a finding is particularly important since X-rays are easily accessible and low cost. These methods not only can diagnose COVID-19 patients from non-COVID pneumonia cases, but can also predict the severity of COVID-19 pneumonia and the risk of short-term mortality. In spite of the low expense of X-ray compared with CT images, the numbers of studies that assessed these two types of imaging using machine/deep learning methods are not meaningfully different. However, few studies have used these methods on both types of imaging (25,29,40). CNN-based methods have achieved accuracy values above 99% in classifying COVID-19 patients from other cases of pneumonia or related disorders, as reported by several independent studies, suggesting these strategies as screening methods for initial evaluation of COVID-19 cases. Although both deep learning and machine learning strategies can be used for the mentioned purpose, they differ in some respects. For instance, deep learning methods usually need a large amount of labeled training data to make a concise conclusion. However, machine learning can apply a small amount of data delivered by users. Moreover, deep learning methods need highperformance hardware. Machine learning, on the other hand, needs features to be precisely branded by users, deep learning generates novel features by itself, thus requires more time to train. Machine learning classifies tasks into small fragments and subsequently combines obtained results into one conclusion, whereas deep learning resolves the problems using end-toend principles.
Several studies have diagnosed COVID-19 patients through the application of machine learning methods rather than using deep learning methods by retrieving the features from the images. These studies have yielded high recognition outcomes and have the advantage of high learning speed (12). Preprocessing is an essential step for reducing the impacts of intensity variations in CT slices and getting rid of noise. Subsequent thresholding and morphological operations have also enhanced the analytical performance. Data augmentation and histogram equalization are among the most applied preprocessing methods.
One of the most promising approaches used in the included studies was transfer learning. Transfer learning is defined as using model knowledge on a huge dataset (which is referred to as the "pre-trained model") and transferring it to use on a new problem. This is very useful in settings like medical imaging, where there is a limited number of labeled data (113). Previous studies showed favorable outcomes of the transfer learning approaches in medical imaging tasks (114,115). Among the included studies, Bridge et al. (25) even reached 100% classification accuracy on COVID-19 using the pre-trained InceptionV3.
The availability of public databases of CT and X-ray images of patients with COVID-19 has facilitated the application of machine learning methods on large quantities of clinical images and execution of training and verification steps. However, since these images have come from various institutes using different scanners, preprocessing of the obtained data is necessary to make them uniform and facilitate further analysis (12). Appraisal of demographic and clinical data of COVID-19 patients and their association with CT/ X-ray images features as well as the accuracy of machine learning prediction methods would provide more valuable information in the stratification of COVID-19 patients. Moreover, one of the major challenges of deep learning models in medical applications is its unexplainable features due to its black-box nature, which should be solved (116). Future studies can focus on approaches that provide interpretation besides black-box predictions.

CONCLUSION
Deep and machine learning methods have high accuracy in the differentiation of COVID-19 from non-COVID-19 pneumonia based on chest images. These techniques have facilitated the automatic evaluation of these images. However, deep learning methods suffer from the absence of transparency and interpretability, as it is not possible to identify the exact imaging feature that has been applied to define the output (13). As no single strategy has the capacity to distinguish all pulmonary disorders based merely on the imaging presentation on chest CT scans, the application of multidisciplinary approaches is suggested for overcoming diagnostic problems (13).

AUTHOR CONTRIBUTIONS
HM-R, MN, and AG-L collected the data and designed the tables. MT and SG-F designed the study, wrote the draft, and revised it. All the authors read the draft and approved the submitted version.