Encyclopedia
Please note this is an old version of this entry, which may differ significantly from the current revision.
Since the COVID-19 pandemic outbreak, over 760 million confirmed cases and over 6.8 million deaths have been reported globally, according to the World Health Organization. While the SARS-CoV-2 virus carried by COVID-19 patients can be identified though the reverse transcription–polymerase chain reaction (RT-PCR) test with high accuracy, clinical misdiagnosis between COVID-19 and pneumonia patients remains a challenge. Therefore, we developed a novel CovC-ReDRNet model to distinguish COVID-19 patients from pneumonia patients as well as normal cases. ResNet-18 was introduced as the backbone model and tailored for the feature representation afterward.
- randomized neural networks
- deep random vector function linking
- convolutional neural networks
1. Introduction
1.1. COVID-19
On 30 January 2020, the World Health Organization (WHO) formally declared the outbreak of COVID-19 and upgraded the pandemic to a public health emergency of international concern (PHEIC). COVID-19, generally identified as coronavirus disease 2019, is a widespread contagious disease caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). According to the epidemiological report from the WHO, over 760 million confirmed cases and over 6.8 million deaths have been reported globally since the beginning of the COVID-19 pandemic until 16 March 2023 [1].
The first people infected by the virus were reported in Wuhan City, Hubei Province, China, and it later spread rapidly across the world [2]. Several studies have confirmed that the COVID-19 virus is primarily transmitted via respiratory droplets and contact routes, resulting in direct human-to-human infection [3,4,5]. Virus transmission happens when people come into close contact (within 1 m) with a confirmed infected person who has respiratory symptoms such as coughing or sneezing, and their exposed mucosae and conjunctiva organ could become a potential receiver of the virus [6]. Common symptoms of COVID-19 include coughing, fever, loss of smell (anosmia), and taste (ageusia). Moreover, long-term consequences occur in post-COVID-19, such as weakness, general malaise, fatigue, cognitive impairment, etc. [7,8].
Diagnosing and detecting coronaviruses significantly contributes to outbreak control and further measures such as isolation and medical treatment. Currently, the mainstream virus detection technology is the reverse transcript–polymerase chain reaction (RT-PCR) test [9,10,11]. According to research from The Lancet Infectious Diseases, when pooled simultaneously, nasal and throat reach a high positive predictive value with an accuracy of 97% [12]. Another comparable detection approach is medical imaging, with different imaging modalities, like computed tomography (CT) and X-ray, being considered the most commonly used technologies [13,14,15]. Although medical imaging has been proven to have limited specificity in identifying COVID-19 (due to overlapping features in chest CT images such as those characterizing adenoviruses, influenza, H1N1, SARS, and MERS) [16], imaging requires more commonly available medical equipment and provides higher sensitivity than the RT-PCR test [17,18]. In addition, medical imaging can be used to confirm diagnostic results when positive-negative RT-PCR test results occur [19,20,21]. Evidently, medical images based on CT and X-ray scans remain highly valuable for COVID-19 disease diagnosis.
1.2. Pneumonia
Pneumonia is an infection that inflames or swells the tissue to create something akin to air sacs (also known as alveoli) in the human respiratory organs, specifically the lungs [22,23,24]. An annually reported 450 million people are infected with pneumonia worldwide, with over 4 million confirmed deaths [25,26]. Hence, it is vital to identify pneumonia at an early stage and further defeat it with prompt medical treatment.
Identifying the responsible pathogen is a crucial part of diagnosing pneumonia, but this is time-consuming and necessitates medical knowledge. Thanks to the rapid development of medical imaging technology, chest CT and X-ray have been proven to be reliable diagnosis approaches since lesions can be directly observed in images. Comparing common pneumonia patients with COVID-19 patients, different features can be captured in medical images. According to Zhao, et al. [27], COVID-19 infections (89.47%) were most commonly distinguished from common pneumonia (6.67%) in patients with ground-glass opacity and multiple mottling in their lung scans. Interestingly, the applicability of AI technology to the task of classifying COVID-19 patients and non-COVID-19 pneumonia patients is theoretical and evidence-based.
Furthermore, the multi-classification task could be more practical compared with binary classification. The reason for this could be that the RT-PCR test has already shown great capability to identify the SARS-CoV-2 virus carried by COVID-19 patients with high accuracy, but distinguishing COVID-19 from other lung diseases still mainly depends on the patient’s medical images. On the other hand, common symptoms between COVID-19 patients and pneumonia patients, such as productive or dry cough, chest pain, fever, and difficulty breathing, confuse the clinic diagnosis. An auto-detection AI system based on chest scans could provide computer-aided detection (CAD) algorithms even if patients have similar clinical symptoms.
The large volume of research on computer-assisted technology significantly contributes to diagnosing and detecting coronaviruses in clinical applications. Common challenges could be described as follows: (a) information loss occurs when deepening the neural network; (b) complex architecture leads to resource waste and time-assuming problems; (c) the network is limited to generalizing different tasks; (d) prediction accuracy remains to be improved.
2. CovC-ReDRNet
Classification tasks in the context of COVID-19 have become increasingly important as the pandemic continues to spread globally. Deep learning models have been applied to various classification problems related to COVID-19, including but not limited to diagnosis, severity assessment, and prognosis prediction. In this section, we highlight some of the recent developments in this field, discuss the challenges and limitations of the existing models, and further provide the motivation for our present research.
One of the earliest and most widely studied classification tasks in COVID-19 is the diagnosis of the disease. A number of studies have proposed deep learning models that can diagnose COVID-19 based on chest X-ray images and CT scans. In 2020, COVID-Net [28] boomed the application of deep learning for detecting COVID-19 cases from chest X-ray images. Additionally, the largest open access benchmark dataset of COVID-19-positive cases was generated, namely COVIDx, which comprises 13,975 chest X-ray images across 13,870 patient cases and is constantly expanding.
Subsequently, COVIDX-Net [29] was proposed to assist radiotherapists in automatically diagnosing COVID-19 based on chest X-ray images. The proposed framework included seven different architectures of deep convolutional neural networks (CNNs). Experimentally, a good performance was achieved by VGG-19 and DenseNet with F1-scores of 89% and 91% for normal and COVID-19 classes, respectively. More recent studies [30,31,32,33,34] supported deep learning approaches to learn discriminative patterns from chest X-ray images and CT scans as well as achieved high accuracy in COVID-19 detection tasks. The contributions and limitations of SOTA methods in the COVID-19 diagnosis task are analyzed in Table 1.
Table 1. The analysis of SOTA methods in COVID-19 diagnosis task.
| Methodology | Tasks | Contributions | Limitations |
|---|---|---|---|
| Wang, et al. [28] | Diagnosis | The study developed COVID-Net as one of the earliest open-source networks for COVID-19 diagnosis. The study assembled one of the largest publicly available datasets of COVID-19-positive cases. |
The study achieved limited model performance with an accuracy of 93.57%. |
| Hemdan, et al. [29] | Diagnosis | The study tested seven different architectures of deep CNN models in COVIDX-Net. | The study merely implemented a binary classification task. The study achieved limited model performance with the highest accuracy being 90%. |
| Narin, et al. [30] | Diagnosis | Three different datasets were tested in the experiments. | The study merely implemented a binary classification task. |
| Rahman, et al. [31] | Diagnosis | Six different architectures of deep CNN models were investigated for multiclassification tasks. | The classification task mainly focused on distinguishing COVID-19 from other brain diseases but not lung disease. |
| Abbas, et al. [32] | Diagnosis | The study addressed the problem of irregularities in annotated data using the DeTraC network. | The classification task achieved limited model performance with the highest accuracy being 93.1%. |
| Zhang, et al. [33] | Diagnosis | The study proposed a lightweight architecture that takes only 1.06 s on average to diagnose a chest CT image. | The study merely implemented a binary classification task. |
| Park, et al. [34] | Diagnosis/Severity Assessment |
The multi-task model is capable of both classification and severity prediction tasks. | The classification task achieved limited model performance with the highest accuracy being 86.8%. |
Another important branch in the COVID-19 classification task is the assessment of disease severity. The severity of COVID-19 can vary greatly from patient to patient, which indicates the importance of identifying patients who are at high risk of developing severe complications. For example, a multi-task vision transformer (ViT) that leverages a low-level chest X-ray feature corpus obtained from a backbone network to diagnose and quantify the severity of COVID-19 was proposed by Park, et al. [34]. The severity quantification performance of the proposed model was evaluated in terms of mean squared error (MSE) with a 95% confidence interval (CI) of 1.441 (0.760–2.122), 1.435 (1.195–1.676), and 1.458 (1.147–1.768) in three external datasets, respectively. Additionally, Goncharov, et al. [35] proposed a CNN-based network that leverages all available labels within a single model, which outperformed existing approaches and achieved a 97% Spearman correlation in severity quantification.
More advanced deep neural networks have been proposed based on various clinical and demographic factors for severity assessment [36,37,38,39]; CNNs and recurrent neural networks in particular have been applied to this task with promising results. The contributions of SOTA methods to the COVID-19 severity assessment task are highlighted in Table 2. Therefore, deep learning methods could be used to determine the prognosis of patients with COVID-19 and further guide clinical decision making.
Table 2. The contributions of SOTA methods to the COVID-19 severity assessment task.
| Methodology | Tasks | Contributions |
|---|---|---|
| Park, et al. [34] | Severity Assessment/Diagnosis | The multi-task model is capable of both classification and severity prediction tasks. |
| Goncharov, et al. [35] | Severity Assessment | The study obtained the most abundant spatial feature patterns, thus improving the quality and accuracy of the model classification significantly. |
| Signoroni, et al. [36] | Severity Assessment | The study proposed an end-to-end multi-network, highlighting its multi-task self-attentive behavior with high robustness for all variabilities from the medical domain. |
| Kollias, et al. [37] | Severity Assessment/Diagnosis | The study created a 3D database, COV19-CT-DB, consisting of chest CT scans from COVID-19 patients. The study developed a hybrid CNN-RNN model for severity classification. |
| Chieregato, et al. [38] | Severity Assessment | In this study, CT patterns were captured using the 3D-CNN model and in turn, the Boruta algorithm was used to select the theoretical values for the SHAP game, where an AUC of 94.9% was achieved. |
| Bougourzi, et al. [39] | Severity Assessment/Diagnosis | The study proposed an Inception-based ensemble architecture for COVID-19 severity assessment, named CNR-IEMN-CSD. The novel network ranked third in the second COV19D competition, demonstrating an improvement of 6.81% over the baseline results. |
A further remarkable application is the prognosis prediction of COVID-19, which refers to the prediction of the outcome of the disease, such as recovery or death. Prognosis prediction is imperative for clinical decision making and resource allocation, as well as for the development of effective treatments. A deep-learning-based study [40] demonstrated its potential to forecast the number of upcoming COVID-19 infections, and could thus significantly contribute to epidemic control. Four standard forecasting models were tested for predicting newly infected cases, deaths, and recoveries in the ten following days. Another study [41] pointed out the importance of prognosis prediction with the aim of triaging patients effectively; thus, mortality of COVID-19 patients was forecasted for one aspect of prognosis. Better performances were obtained using LASSO and linear SVM, with sensitivities of 90.7% and 92.0%, specificities of 91.4% and 91.8%, and area under the receiver operating characteristics curves (AUCs) of 96.3% and 96.2%, respectively.
More recently, several studies proposed various deep learning architectures for prognosis prediction [42,43,44,45], such as feedforward neural networks (FFNNs) and gradient boosting machines (GBMs), which showed that deep learning models could provide reliable predictions of patient condition, and further provide a deep understating of virology as well as aid in disease control. The contributions of SOTA methods to the COVID-19 prognosis task are highlighted in Table 3.
Table 3. The contributions of SOTA methods in the COVID-19 prognosis task.
| Methodology | Task | Contribution |
|---|---|---|
| Rustam, et al. [40] | Prognosis | The study contained four traditional prediction machine learning models, that is, linear regression, least absolute shrinkage and selection operator (LASSO), support vector machine (SVM), and exponential smoothing (ES), with the aim of predicting the risk level of COVID-19 spread. |
| An, et al. [41] | Prognosis | The study investigated LASSO, SVM, random forest (RF), and K-nearest neighbor (KNN) in order to predict mortality and thus achieve accurate prognostic predictions to triage patients effectively. |
| Khan, et al. [42] | Prognosis | The study examined RF, KNN, decision tree (DT), logistic regression (LR), extreme gradient boosting (XGBoost), and deep learning networks to forecast mortality in COVID-19 cases. The models were trained on confirmed COVID-19 patients from 146 countries. |
| Ikemura, et al. [43] | Prognosis | The study developed the autoML framework to investigate 20 machine learning models to generate the best ensemble model based on 48 variables. In particular, an AUPRC of 80.7% was recorded. The study identified critical variables associated with mortality and accurately forecasted the survival of COVID-19 patients. |
| Elshennawy, et al. [44] | Prognosis | The study developed three architectures, a basic CNN-based (named CV-CNN) model, a hybrid CNN combining a long short-term memory (LSTM) mechanism with a CNN model (named CV-LSTM + CNN), and a hybrid model trained using transformed images (named IMG-CNN). In particular, the average accuracy of the IMG-CNN prognostic model reached 94.14%. |
As mentioned above, deep learning technologies are effective in solving various classification tasks related to COVID-19, including diagnosis, severity assessment, and prognosis prediction. However, there have been a limited number of multi-category classification tasks developed. A multi-category classification task based on deep learning algorithms could be used to accurately diagnose COVID-19 and distinguish it from other respiratory illnesses such as the flu, pneumonia, and other viral infections. It is of considerable importance that the symptoms of COVID-19 are similar to those of many other respiratory illnesses, and misdiagnosis can cause serious consequences for both the patient and public health.
Some three-category classification frameworks that distinguish COVID-19 patients from pneumonia patients and normal cases have been proposed in recent years. Hussain, et al. [46] proposed a CNN-based model dedicated to COVID-19 diagnosis and classification, named CoroDet. A novel database, the COVID-R dataset, was constructed by merging and revising eight COVID-19 open sources, containing 7390 pulmonary images from 2843 COVID-19 patients, 3108 normal cases, and 1439 pneumonia patients. In their three-category classification experiments, the presence of the pulmonary lesion feature of COVID-19 disease in X-ray images was used to differentiate COVID-19 infection from non-COVID-19 pneumonia. CoreDet measured through sensitivity, specificity, precision, recall, F1-score, and accuracy, achieving a good performance based on the average of five-fold cross-validation, that is, 92.76%, 94.56%, 94.04%, 92.50%, 91.32%, and 94.20%, respectively.
Xu, et al. [47] proposed a novel approach for COVID-19 screening, distinguishing COVID-19 from other types of viral pneumonia, especially influenza-A viral pneumonia (IAVP), based on pulmonary CT images. A total of 618 CT images were obtained from three top hospitals in China, including 219 COVID-19 cases, 224 IAVP cases, and 175 normal cases. An advanced model was developed based on the classic ResNet-18 with a location attention mechanism, achieving an overall accuracy of 86.7%. The three different measurements considered, recall, precision, and F1-score, were 86.7%, 81.3%, and 83.9% in the COVID-19 group; 83.3%, 86.2%, and 84.7% in the IAVP group; and 90.0%, 93.1%, and 91.5% in the normal group, respectively.
This entry is adapted from the peer-reviewed paper 10.3390/make5030037
This entry is offline, you can click here to edit this entry!
