This paper is in the following e-collection/theme issue:

Formative Evaluation of Digital Health Interventions (2197) JMIR Theme Issue: COVID-19 Special Issue (2388) Machine Learning (1375) Artificial Intelligence (1191) Decision Support for Health Professionals (1131) Clinical Information and Decision Making (1319) Imaging Informatics (184)

Published on in Vol 7 (2023)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/42324, first published .
Classifying COVID-19 Patients From Chest X-ray Images Using Hybrid Machine Learning Techniques: Development and Evaluation

Classifying COVID-19 Patients From Chest X-ray Images Using Hybrid Machine Learning Techniques: Development and Evaluation

Classifying COVID-19 Patients From Chest X-ray Images Using Hybrid Machine Learning Techniques: Development and Evaluation

Authors of this article:

Thanakorn Phumkuea1 Author Orcid Image ;   Thakerng Wongsirichot2 Author Orcid Image ;   Kasikrit Damkliang2 Author Orcid Image ;   Asma Navasakulpong3 Author Orcid Image
Article Authors Cited by Tweetations (3) Metrics

    Original Paper

    1College of Digital Science, Prince of Songkla University, Songkhla, Thailand

    2Division of Computational Science, Faculty of Science, Prince of Songkla University, Songkhla, Thailand

    3Division of Respiratory and Respiratory Critical Care Medicine, Faculty of Medicine, Prince of Songkla University, Songkhla, Thailand

    *these authors contributed equally

    Corresponding Author:

    Thakerng Wongsirichot, PhD

    Division of Computational Science, Faculty of Science

    Prince of Songkla University

    15 Kanjanavanich Road, Hat Yai

    Songkhla, 90110

    Thailand

    Phone: 66 846414784

    Email: thakerng.w@psu.ac.th


    Background: The COVID-19 pandemic has raised global concern, with moderate to severe cases displaying lung inflammation and respiratory failure. Chest x-ray (CXR) imaging is crucial for diagnosis and is usually interpreted by experienced medical specialists. Machine learning has been applied with acceptable accuracy, but computational efficiency has received less attention.

    Objective: We introduced a novel hybrid machine learning model to accurately classify COVID-19, non-COVID-19, and healthy patients from CXR images with reduced computational time and promising results. Our proposed model was thoroughly evaluated and compared with existing models.

    Methods: A retrospective study was conducted to analyze 5 public data sets containing 4200 CXR images using machine learning techniques including decision trees, support vector machines, and neural networks. The images were preprocessed to undergo image segmentation, enhancement, and feature extraction. The best performing machine learning technique was selected and combined into a multilayer hybrid classification model for COVID-19 (MLHC-COVID-19). The model consisted of 2 layers. The first layer was designed to differentiate healthy individuals from infected patients, while the second layer aimed to classify COVID-19 and non-COVID-19 patients.

    Results: The MLHC-COVID-19 model was trained and evaluated on unseen COVID-19 CXR images, achieving reasonably high accuracy and F measures of 0.962 and 0.962, respectively. These results show the effectiveness of the MLHC-COVID-19 in classifying COVID-19 CXR images, with improved accuracy and a reduction in interpretation time. The model was also embedded into a web-based MLHC-COVID-19 computer-aided diagnosis system, which was made publicly available.

    Conclusions: The study found that the MLHC-COVID-19 model effectively differentiated CXR images of COVID-19 patients from those of healthy and non-COVID-19 individuals. It outperformed other state-of-the-art deep learning techniques and showed promising results. These results suggest that the MLHC-COVID-19 model could have been instrumental in early detection and diagnosis of COVID-19 patients, thus playing a significant role in controlling and managing the pandemic. Although the pandemic has slowed down, this model can be adapted and utilized for future similar situations. The model was also integrated into a publicly accessible web-based computer-aided diagnosis system.

    JMIR Form Res 2023;7:e42324

    doi:10.2196/42324

    Keywords

    COVID-19; machine learning; medical informatics; coronavirus; diagnosis; model; detection; healthy; unhealthy; public; usage; data; database; accuracy; development; x-ray; imaging 


    COVID-19 has become a widespread pandemic causing high levels of infection and mortality. The first cases were reported in December 2019 and rapidly spread worldwide, leading to its declaration as a severe disease by the World Health Organization (WHO) in May 2020 [1]. Recently, more than 190 million confirmed cases have been reported, with 4 million fatalities worldwide [2]. The virus responsible for COVID-19, a severe acute respiratory syndrome (SARS), was formally named SARS-CoV-2 by the International Committee on Taxonomy of Viruses [3]. An initial study showed that the virus originated from bats and was transmitted to humans by unknown intermediate animals [4,5]. COVID-19 has been divided into 5 clinical stages based on its characteristics: asymptomatic, mild clinical symptoms, moderate clinical features, severe symptoms, and critical cases [6]. The majority of COVID-19 patients have fever, fatigue, cough, shortness of breath, myalgia, and dyspnea [7,8]. However, patients may have asymptomatic COVID-19 disease [9]. Reverse transcription polymerase chain reaction (RT-PCR) performed on throat swab samples is the gold standard for COVID-19 diagnosis [10]. However, RT-PCR results require a considerable amount of time to become available [11]. Rapid diagnostic methods (such as the rapid antigen and antibody tests) are available; however, they cannot substitute for RT-PCR [12-14]. Generally, a chest x-ray (CXR) is prescribed for a high-risk patient—old age, high blood pressure, and chronic respiratory disease—who may be classified as a patient under inspection [15]. Thereafter, a CXR image is examined by medical doctors or specialists for lung infection. A challenging situation occurs because the lung infection may be due not only to SARS-CoV-2 but also to other viruses and bacteria [16]. During the widespread COVID-19 pandemic, many areas with a high level of infection experienced a shortage of physicians.

    Recent work explored deep learning and medical imaging techniques to diagnose CXR film of COVID-19 for early detection purposes [17-20]. In [21], the researchers presented a series of steps for classifying COVID-19 and other lung diseases based on 2 data sets: a CXR data set of 1926 images and a computed tomography (CT) scan data set of 2482 images. The results from the CXR data set have an accuracy of 0.993 and an F1-score of 0.931. Furthermore, the results from the CT scan data set have an accuracy of 0.932 and an F1-score of 0.921. The experiment used Raspberry Pi Linux and Python code to perform a sequential feature selector. A similar work, documented in [22], implemented VGG16 and Xception to distinguish COVID-19 infections from noninfected cases. They developed 2 models using 1037 CXR images (402 COVID-19 images, 400 normal images, 200 pneumonia images, and 35 images without COVID-19 or pneumonia infection). Regarding this study, each convolutional neural network (CNN) architecture was subjected to 10 rounds of experimentation for model evaluation purposes. The highest accuracy rate achieved was 0.970 for the VGG16 and 0.984 for Xception. An Xception model [23] was proposed to classify COVID-19 CXR images, with 0.896 and 0.950 accuracies for 4 and 3 classes of classification, respectively. The data set consisted of 284 COVID-19, 327 viral pneumonia, 330 bacterial pneumonia, and 310 normal CXR images. In [24], researchers proposed a classification model to classify COVID-19 from CXR images. The study comprised 4290 pneumonia, 1583 normal, and 76 COVID-19 CXR images. Furthermore, pneumonia and normal classes were partially used and split to balance the data set. The researchers focused on the use of a data preprocessing step. An image augmentation technique was used with COVID-19 CXR images. The fine-tuning of a pretrained model, SqueezNet, used Bayesian optimization. The proposed model obtained a test accuracy and F-measure of 0.983 and 0.983, respectively. Four transfer learning techniques, ResNet18, ResNet50, SqueezeNet, and DenseNet-121, were used to identify COVID-19 on CXR images in [25]. The COVID-19 CXR data set was divided into 2084 training and 3100 test images. The results showed that SqueezeNet achieved 0.929 specificity and 0.98 sensitivity. However, there was no presence of accuracy of the selected transfer learning techniques. A similar study was conducted in [26]. Pretrained AlexNet, GoogLeNet, and SqueezeNet models were fine-tuned to classify lung infections from CXR images. They used 6 CXR image data sets that were collected from several public databases. The classification results of the pretrained models were promising. Another research work [27] used the transfer learning technique with GoogLeNet, ResNet-18, and DenseNet-12 to perform binary classification of normal and pneumonia images from 2 publicly available CXR image data sets; a 5-fold cross-validation technique was used to evaluate the models. The experimental results showed accuracy and sensitivity of 0.988 and 0.988, respectively, for the first data set and 0.869 and 0.870, respectively, for the second data set. In [28], a CNN-based architecture was proposed to delineate CXR images into 3 categories: healthy, pneumonia, and COVID-19. The data sets were collected from 6 public databases, including 10,451 healthy, 573 COVID-19, and 11,673 pneumonia images. The proposed model achieved an accuracy of 0.912 in the prediction of the 3 classes (healthy, pneumonia, and COVID-19) and an accuracy of 0.982 in the prediction of the 2 classes (COVID-19 or pneumonia). In [29], a new end-to-end trained CNN model was proposed with deep feature extraction and fine-tuning of a pretrained CNN. The proposed model consisted of 3 deep learning approaches for COVID-19 detection. The data set used in this study contained 180 COVID-19 and 200 healthy CXR images. The selected features extracted from the ResNet50 model and support vector machine (SVM) classifiers outperformed other approaches, with a classification accuracy of 0.947. In [30], 5 different pretrained deep models (decision tree [DT], random forest, AdaBoost, bagging, and SVM) were used to classify COVID-19 CXR. With this experiment, the data set contained 1102 CXR images: 565 normal and 537 COVID-19. The data set was divided into training and test sets at a ratio of 70:30. The results showed that the Xception combined with the SVM classifier achieved the best classification result, with accuracy, sensitivity, and specificity of 0.993, 0.992, and 0.993, respectively. This study was exceptional by proposing combinations of traditional machine learning and deep learning techniques. A hybrid ensemble model using MobileNet and Inception V3 was proposed in [31]; 4-fold cross-validation was performed to evaluate the model using a data set of 1050 normal, 1050 viral pneumonia, 1050 bacterial pneumonia, and 1050 COVID-19 CXR images. The classification results for diagnosing COVID-19 from CXR achieved accuracy, precision, and specificity of 0.942, 0.899, and 0.883, respectively. Table 1 shows a summary of related studies on COVID-19 CXR images.

    Table 1. Related studies of COVID-19 chest x-ray images.
    MethodData setBalanced data setClasses, nCross-validationReference
    Sequential feature selectionCOVID-19=212, non-COVID-19=1696No25-fold[21]
    Xception and VGG16COVID-19=402, pneumonia=200, normal=400, without COVID-19 or pneumonia=35No2No[22]
    CoroNet (Xception)COVID-19=284, bacterial pneumonia=330, viral pneumonia=327, normal=310No3 and 44-fold[23]
    Bayes-SqueezeNetCOVID-19=1979, pneumonia=3895, normal=3111No3No[24]
    Different pretrained CNNa modelCOVID-19=184, normal=5000No25-fold[25]
    Different pretrained CNN model6 different COVID-19 data setsNo2 and 3No[26]
    GoogLeNet, ResNet-18, and DenseNet-1212 different pneumonia data setsNo25-fold[27]
    Customized CNN modelCOVID-19=573, normal=10,546, pneumonia=11,673No2No[28]
    ResNet50 for Features extraction + SVMbCOVID-19=180, normal=200No2No[29]
    Xception + SVMCOVID-19=537, normal=565No210-fold[30]
    Hybrid ensemble modelCOVID-19=1050, viral pneumonia=1050, bacterial pneumonia=1050, normal=1050Yes44-fold[31]
    MLHC-COVID-19COVID-19=1050, non-COVID-19=2100, healthy=1050Yes310-foldOur proposed method

    aCNN: convolutional neural network.

    bSVM: support vector machine.

    According to the literature review, most of the reported studies used deep learning and a pretrained model with data augmentation for COVID-19 detection from CXR images. However, there are several limitations that remain to be addressed. These limitations include (1) the use of an imbalanced data set, (2) the use of deep learning techniques that require significant computational resources, and (3) longer time consumption of the training model. The main contributions of our study, with the aim of addressing these issues, are as follows:

    1. We proposed a multilayer hybrid classification model for COVID-19 (MLHC-COVID-19) detection. Our model integrates several machine learning techniques applied to a large CXR image data set of 3 different categories: healthy, non-COVID-19 (viral and bacterial pneumonia), and COVID-19. The MLHC-COVID-19 model was evaluated through a 10-fold cross-validation process to assess its classification performance.
    2. The MLHC-COVID-19 is composed of 2 layers of binary classification. The first layer acted as a screening mechanism, directing unhealthy CXR images to the second layer for further classification into COVID-19 and non-COVID-19 images. This model has been thoroughly compared with other preprocessing techniques and methods to assess its effectiveness.
    3. The first layer of the MLHC-COVID-19 uses the highest performance model between DTs, SVMs, and neural networks (NNs) to differentiate between healthy and unhealthy CXR images. The second layer uses the most effective model between the same 3 techniques to distinguish between COVID-19 and non-COVID-19 CXR images. The model performance was evaluated through 10-fold cross-validation.
    4. We developed a web-based application for the MLHC-COVID-19 model, allowing for practical use in diagnosing COVID-19 from CXR images. The web-based application can be accessed at [32].

    This paper is organized as follows: The proposed method (MLHC-COVID-19) is discussed in the Methods section. The Results section presents the in-depth experimental results, compared with deep learning–based methods. The conclusion of the study and suggestions for future research are discussed in the Discussion section.


    Data Set

    In a real-world situation, lung infection can be due to many factors. The distinction of our work lies in the combination of various types of lung infection and healthy CXR images from different data sources. The target classes were healthy, COVID-19, viral pneumonia, and bacterial pneumonia. We used 5 data sets of CXR images that were made publicly available by Mendeley [33], the Italian Society of Medical and Interventional Radiology [34], GitHub [35], Radiopaedia [36], and Kaggle [37]. The total number of images used in this study was 4200. We utilized an annotated data set of CXR images consisting of 3 distinct classes: (1) healthy individuals, (2) individuals infected with non-COVID-19 diseases (viral and bacterial pneumonia), and (3) individuals infected with COVID-19. The COVID-19 class was comprised of 1050 CXR images, of which 912 images were obtained from the first 4 sources and 138 images were obtained from the last source. The non-COVID-19 and healthy classes were comprised of a total of 3150 CXR images, with 1050 images being healthy and 2100 images being non-COVID-19. The annotation of each image was performed by skilled medical specialists. Table 2 shows a summary of the CXR images data set used in this study, and samples of the CXR images in the combined data set are shown in Figure 1.

    Table 2. Chest x-ray image data set.
    Open source/classesNumber of images, n
    Kaggle [37]

    		Healthy1050

    		Non-COVID-192100
    Mendeley data set [33]

    		COVID-19912
    SIRMa [34], GitHub [35], and Radiopaedia [36]

    		COVID-19138

    aSIRM: Italian Society of Medical and Interventional Radiology.

    ‎
    Figure 1. Samples of the chest x-ray images in the combined data set: (A) heathy images, (B) viral pneumonia images, (C) bacterial pneumonia images, (D) COVID-19 images.

    Overview

    The general overview of the proposed MLHC-COVID-19 for identifying CXR images of COVID-19 infection is shown in Figure 2. The block diagram depicts vital processes that were embedded in the model. The following subsections will discuss each of the processes.

    ‎
    Figure 2. Block diagram of the proposed multilayer hybrid classification model (MLHC)-COVID-19.

    Image Preprocessing

    In this study, image preprocessing consisted of 3 steps (Figure 3): image segmentation, image analysis, and feature extraction. Considering the image segmentation, we designed 3 substeps: removal of the partial black background, image resizing, and grayscale conversion. The combined data set had different formats and a partial black background in the CXR images. This was attributed to several factors including the positioning during the filming process and the type of x-ray machines utilized. The black background affects the classification performance. We programmatically removed the black background from the CXR image. Figure 4 shows an example of a completed programmatic removal of a black background. Furthermore, the CXR images were of different dimensions; therefore, all the images were resized to an identical dimension of 500 × 500 pixels. The last step of the image segmentation was the conversion of the images to grayscale to reduce the image features. Reduction of the image features is known to improve the classification result and mimic the complexity of the algorithm [38]. The grayscale formula is given by equation 1.

    Gintensity(x,y)=0.2989 × f(x,y,R) + 0.5871 × f(x,y,G) + 0.1140 × f(x,y,B) (1)

    where, if Gintensity(x,y) is an image with grayscale, then f(x,y,R) is a pixel value in the (x,y) coordinates of the red channel, f(x,y,G) is a pixel value in the (x,y) coordinates of the green channel, and f(x,y,B) is a pixel value in the (x,y) coordinates of the blue channel [39].

    The second step in the image preprocessing was image enhancement. We utilized the power law transformation to adjust the brightness of the CXR images. We applied a γ value of 0.5. The power-law transformation formula is shown in equation 2 [40,41]. In addition, we selected the 2-dimensional Gaussian filter technique to reduce the Gaussian and salt-and-pepper noises [42]. The Gaussian filter technique is given by equation 3. Figure 5 shows the processed CXR image after the image enhancement.

    s=c × rγ(2)

    where s is the output pixel value, c is a value of the normalized image, γ is the gamma value, and r is the input pixel value.

    where ∂2 is the variance of the Gaussian filter with 3 x 3 kernel size and x and y are the horizontal and vertical axes, respectively, of the kernel size [43].

    The last step in image preprocessing is feature extraction. We performed histogram analysis and L2-normalization. Histogram analysis reduces the image features by retrieving vital image statistics. Specifically, low-intensity values dictate an image in a dark tone and vice versa [44]. Upon completion of the histogram analysis, each feature in the histogram has different scales. Subsequently, to achieve the same standard scale, we performed L2-normalization. L2-normalization or the Euclidean norm normalizes the features of the histogram into the same scale using equation 4. Figure 6 shows the output of feature extraction.

    where X is a feature of the histogram.

    ‎
    Figure 3. Image preprocessing.
    ‎
    Figure 4. (A) An original chest x-ray (CXR) image with a partial black background and (B) the CXR image after the removal of the partial black background.
    ‎
    Figure 5. Image after the (A) image segmentation, (B) power law transformation, (C) Gaussian filter.
    ‎
    Figure 6. The final output of the image preprocessing step.

    Multilayer Hybrid Classification Model

    First, the original MLHC was developed to automatically classify the multisleep stages. Each layer in the MLHC is a binary classification model that uses different machine learning techniques [45]. The MLHC model was a stacking-based machine learning model that utilized multiple models to improve its performance. “Multilayer” refers to the stacking of multiple models, while “hybrid” signifies the combination of different model types, such as DTs, SVMs, and NNs. The primary objective of the MLHC was to diagnose COVID-19 from CXR images. By stacking various models, the MLHC aimed to take advantage of the strengths of each model and enhance the accuracy and robustness of the system in diagnosing COVID-19 patients.

    The MLHC-COVID-19 model in the experiment was designed using a 2-step approach. The first step aimed to differentiate between healthy and unhealthy individuals, where the latter included those infected with viral, bacterial, or COVID-19 infection. Once the differentiation was made, the infected individuals were then directed to the second step, which was to classify them further into either COVID-19 or non-COVID-19 cases (consisting of viral and bacterial infections). Figures 7 and 8 show the MLHC-COVID-19.

    Regarding our experiment, we evaluated 3 machine learning techniques: DT, SVM, and NN. These selected techniques were candidates for embedding into MLHC-COVID19. DT is a well-known supervised machine learning method. Each node represents a condition for decision on data classification, in which various branches of trees represent results from testing and the leaves of the DT represent the classification [46,47]. DT is one of the simplest techniques to understand and is suitable for classification tasks [48]. SVM is a supervised machine learning method. It exhibits promising performance in statistical classifications [49]. It distinguishes data by finding hyperplanes as separators. The process of identifying hyperplanes is iteratively toward the best line during the training [50]. We selected the radial basis function as the kernel function [51]. Finally, NNs are types of mathematical models for processing data with connected computation nodes that mimic the functions of biological NNs [52]. It builds complex models between the inputs and outputs with high efficiency [53]. We designed 4 fully connected layers (dense layers) and 2 dropout layers. The input to the first dense layer consisted of 256 histograms. The first and second dense layers used 128 neurons, and the third dense layer used 32 neurons. The last dense layer was fed into the softmax classifier [54]. In addition, the dropout layers, with a ratio of 0.2, intervened between the dense layers [55].

    ‎
    Figure 7. Pseudocode of the multilayer hybrid classification model (MLHC)-COVID-19.
    ‎
    Figure 8. Flowchart of multilayer hybrid classification model (MLHC)-COVID-19.

    Performance Evaluation

    We divided the data set into a training set and a testing set at a ratio of 80:20. The training set was utilized to train the MLHC-COVID-19 model, and its performance was evaluated through 10-fold cross-validation. The testing set, representing unseen data, was utilized to select the optimal model among the models generated in the training phase. The performance measures were accuracy, sensitivity, specificity, precision, F measure, and area under the curve (AUC) [56,57]. The performance measure formulae are given in equations 5-10, and the measured performances included the true positive (TP), true negative (TN), false positive (FP), and false negative (FN). In addition, the classification performance was evaluated using the area under the receiver operating characteristic curve (AUC of the ROC). We used the Python module 3.8.3 known as Scikit-learn for machine learning algorithms [58] with an NVIDIA GeForce-960 M GPU, 4 GB GDDR5 onboard memory with Intel Core i7-6700 HQ (2.60 GHz, 6 MB L3 cache, approximately 3.50 GHz), 8 GB DDR4 RAM, and 1 TB hard drive.


    Considering the MLHC-COVID-19 design, the distribution of data in each layer was imbalanced, potentially causing bias in the evaluation. To solve the imbalance issue, we used the synthetic minority oversampling technique (SMOTE) to balance the classes [59]. The SMOTE synthesized new data from the existing data using k-nearest neighbors and inserted them into the original data set [60]. The results of using SMOTE are presented in Table 3, which shows the number of images in each class of the original and the SMOTE-augmented data sets based on the MLHC-COVID-19 design. For instance, the original data set consisted of 840 images in the healthy class and 2520 images in the unhealthy class, resulting in the healthy class being the minority class in Layer I. To balance both classes, 1680 synthetic healthy instances were generated using SMOTE.

    We investigated 3 machine learning techniques: DT, SVM, and NNs. The best classification model for each layer was selected for the MLHC-COVID-19. Table 4 shows the classification results for training and testing. During training, 10-fold cross-validation was performed, and the results are shown as the average (SD). First, the NNs achieved the highest accuracy at 0.983 (SD 0.013) during the training, followed by SVM and DT in Layer I. In addition, the NNs achieved the highest position with a mean sensitivity of 0.987 (SD 0.012), specificity of 0.980 (SD 0.022), precision of 0.979 (SD 0.023), F measure of 0.983 (SD 0.013), and AUC of 0.995 (SD 0.005). Considering the time, the DT minimized the time spent in Layer I. Considering Layer II, the SVM had the highest performance metric in all the aspects during the training (mean AUC 0.984, SD 0.004). The second best was the NN, followed by the DT.

    Table 3. Comparison of the number of instances of the original and applied synthetic minority oversampling technique (SMOTE) data sets.
    MLHC-COVID-19a classImages in the training set (80%), n


    		OriginalSMOTE
    Layer I

    		Healthy8402520
    Unhealthy25202520
    Layer II

    		COVID-198401680
    Non-COVID-1916801680

    aMLHC: multilayer hybrid classification model for COVID-19.

    Table 4. 10-fold cross-validation results.
    Layer and performance metricTrain (10-fold cross-validation)


    		DTa, mean (SD)SVMb, mean (SD)NNc, mean (SD)
    Layer I (healthy, unhealthy)

    		Accuracy0.863 (0.018)0.962 (0.006)0.988 (0.010)d

    		Sensitivity0.719 (0.048)0.972 (0.009)0.992 (0.007)d

    		Specificity0.912 (0.009)0.953 (0.012)0.985 (0.015)d

    		Precision0.728 (0.042)0.954 (0.011)0.985 (0.014)d

    		F measure0.723 (0.040)0.963 (0.005)0.989 (0.009)d

    		AUCe0.819 (0.025)0.962 (0.006)0.995 (0.005)d

    		Training time (sec)1.108 (0.066)d1.201 (0.123)24.888 (1.277)
    Layer II (COVID-19, non-COVID-19)

    		Accuracy0.936 (0.012)0.985 (0.004)d0.979 (0.005)

    		Sensitivity0.950 (0.012)0.981 (0.007)d0.977 (0.007)

    		Specificity0.908 (0.026)0.987 (0.005)d0.981 (0.015)

    		Precision0.954 (0.011)0.987 (0.005)d0.981 (0.015)

    		F measure0.952 (0.009)0.984 (0.004)d0.979 (0.005)

    		AUC0.929 (0.015)0.984 (0.004)d0.979 (0.005)

    		Training time (sec)0.749 (0.070)0.275 (0.020)d18.183 (0.493)

    aDT: decision tree.

    bSVM: support vector machine.

    cNN: neural network.

    dBest classification results.

    eAUC: area under the curve.

    Figure 9 shows a graphical representation of the averaged ROC of the 0-fold cross-validation of the training. The NNs achieved the best classification result in Layer I, and SVM outperformed all the other models in Layer II.

    Considering the 10-fold cross-validation, the proposed MLHC-COVID-19 model consisted of 2 layers, in which NNs were utilized in the first layer and SVM was utilized in the second layer. The first layer was designed to differentiate between healthy and unhealthy based on CXR images. The second layer then classified the infected individuals into either COVID-19 or non-COVID-19. For the testing, we separated 20% of the data set as mentioned in the previous section. The classification results in the unseen test also confirmed that the NNs and SVM are the best classifiers in Layers I and II, respectively. However, to address the issue of class imbalance in the unseen test set, we applied the SMOTE to the data as shown in Table 5.

    The MLHC-COVID-19 model demonstrated superior performance compared with other techniques in the unseen test results. As presented in Table 6, the accuracy and F measure of the MLHC-COVID-19 model in the unseen test set were 0.962 and 0.962, respectively. Figure 10 displays the full confusion matrix of the MLHC-COVID-19 in the unseen test. The MLHC-COVID-19 correctly identified 415 cases of COVID-19. There were 2 COVID-19 cases that were misclassified as non-COVID-19, and 3 COVID-19 cases were incorrectly classified. Conversely, there were 8 cases (7 non-COVID-19 and 1 healthy) that were incorrectly classified as COVID-19 cases. For the non-COVID-19 class, 7 cases were misclassified as COVID-19, while 400 cases were correctly classified, and 13 cases were misclassified as healthy. For the healthy class, only 1 case was misclassified as COVID-19, 22 cases were misclassified as non-COVID-19, and 397 cases were correctly classified.

    ‎
    Figure 9. Receiver operating characteristic (ROC) in (A) Layer I and (B) Layer II. ANN: artificial neural network; AUC: area under the curve; DT: decision tree; SVM: support vector machine.
    Table 5. Results of synthetic minority oversampling technique (SMOTE) on the unseen testing set.
    ClassImages in the testing set (20%), n
    OriginalSMOTE
    COVID-19210420
    Non-COVID-19420420
    Healthy210420
    Table 6. Classification results from the multilayer hybrid classification model (MLHC)-COVID-19.
    ClassPrecisionSensitivityF measureAccuracy
    COVID-190.9810.9880.985—a
    Non-COVID-190.9430.9520.948—
    Healthy0.9610.9450.953—
    Macro-average0.9620.9620.9620.962

    aNot assessed.

    ‎
    Figure 10. Confusion matrix of the multilayer hybrid classification model (MLHC)-COVID-19.

    Principal Findings

    Our study showed that the MLHC-COVID-19 model achieved promising results in terms of accuracy, sensitivity, and precision for classifying healthy, non-COVID-19, and COVID-19 images. Image preprocessing played a crucial role in obtaining meaningful information and accurate classification by eliminating noisy or distorted pixels from each image. This allowed the classification model to diagnose infections from CXR images effectively. Additionally, histogram analysis of CXR images and L2-normalization reduced the training time of the model to less than 30 seconds. The data set was also divided based on a newly proposed multilayer design.

    Comparison With Prior Work

    Diagnosis of COVID-19 was based on analyzing or classifying CXR images. Recent studies showed various preprocessing techniques, feature extraction methods, and classification approaches. Two main deep learning techniques, the customized CNN and ensemble learning methods, were extensively used in most recent research. Due to the complexity of the deep learning techniques, they require extensive computer processing performance. We aimed to develop the MLHC-COVID-19 model to reduce the processing time compared with other deep learning techniques while still maintaining acceptable performance. Our objective was to achieve a reliable COVID-19 diagnosis from CXR images and reduce the likelihood of misdiagnosis. The model achieved an accuracy of over 95%, which was higher than other similar work. These results are detailed in Table 7, which compares our model with those in other studies. Our MLHC-COVID-19 model outperformed the hybrid ensemble model, which achieved an accuracy of 0.942, sensitivity of 0.884, precision of 0.899, and F measure of 0.886 on the same data set in [31]. Our proposed model achieved a higher accuracy of 0.962 and F measure of 0.962.

    Table 7. Comparison of the COVID-19 classification results.
    Number of classesHybrid or ensemble modelAccuracySensitivitySpecificityPrecisionF measureAUCaStudy
    2No0.9900.9680.9910.9090.938—b[21]
    2Yes0.9690.9540.9760.9540.954—[22]
    3No0.9500.9690.9750.9500.956—[23]
    3Yes0.983—0.991—0.983—[24]
    2Yes—0.9800.900———[25]
    3Yes0.9750.9540.97760.9850.9690.999[26]
    3No0.912—————[28]
    2Yes0.9470.9110.9800.9760.943—[29]
    2Yes0.9930.9900.9940.9930.9930.993[30]
    4Yes0.9420.884—0.8990.886—[31]
    3No0.9620.962—0.9620.962—MLHC-COVID-19c

    aAUC: area under the curve.

    bNot assessed.

    cMLHC: multilayer hybrid classification model for COVID-19.

    Limitations

    The MLHC-COVID-19 model has some limitations that can be improved upon. Currently, the model is designed for global classification, meaning it cannot identify specific abnormal regions within CXR images. We plan to explore other data preprocessing techniques beyond histogram analysis that may lead to better classification results. Additionally, there are still other techniques available for COVID-19 classification that we aim to explore. Last, our goal is to enhance the MLHC-COVID-19 model by adding another layer that can classify viral pneumonia and bacterial pneumonia, resulting in a 4-class classification system.

    Conclusion

    We developed the MLHC-COVID-19 to diagnose COVID-19 from CXR images. Our model combines multiple individual models to leverage the strengths of each and improve the overall accuracy and robustness of the system for diagnosing COVID-19. The MLHC-COVID-19 model was designed to differentiate COVID-19–infected CXR images from healthy and non-COVID-19 images, such as those affected by viral or bacterial pneumonia. The model was thoroughly evaluated and compared with other preprocessing techniques and methods to assess its effectiveness. The findings of this study were acceptable when compared with the other techniques. Considering the current situation, the computer-aided diagnosis tool must be easily accessible; therefore, a web-based solution is also feasible. In terms of the practical implication, we developed a prototype for a web-based computer-aided diagnosis tool [32] as an alternative to bring the MLHC-COVID-19 technique into the clinical setting as a useful tool for supporting radiologists in improving COVID-19 accuracy. The main system of this tool allows users to upload CXR images, which are then diagnosed by the MLHC-COVID-19. The tool will evaluate whether the image is healthy, non-COVID-19, or COVID-19.

    Acknowledgments

    This work was supported by the Digital Science for Economy, Society, Human Resources Innovative Development and Environment project funded by Reinventing Universities & Research Institutes under grant number 2046735, Ministry of Higher Education, Science, Research and Innovation, Thailand.

    Data Availability

    The data are available at [61].

    Authors' Contributions

    TP, TW, KD, and AN conceptualized the study and reviewed and edited the manuscript. TP and TW performed the formal analysis and wrote the original draft of the manuscript. TP was responsible for the software and visualization. TP, TW, and KD designed the methodology. TW supervised the study. TP, TW, and AN performed the validation.

    Conflicts of Interest

    None declared.

    1. Yuki K, Fujiogi M, Koutsogiannaki S. COVID-19 pathophysiology: A review. Clin Immunol 2020 Jun;215:108427 [FREE Full text] [CrossRef] [Medline]
    2. COVID-19 Weekly Epidemiological Update. World Health Organization. 2021 Jan 10.   URL: https:/​/www.​who.int/​docs/​default-source/​coronaviruse/​situation-reports/​weekly_epidemiological_update_22.​pdf [accessed 2023-02-18]
    3. Coronaviridae Study Group of the International Committee on Taxonomy of Viruses. The species severe acute respiratory syndrome-related coronavirus: classifying 2019-nCoV and naming it SARS-CoV-2. Nat Microbiol 2020 Apr;5(4):536-544 [FREE Full text] [CrossRef] [Medline]
    4. Singhal T. A review of coronavirus disease-2019 (COVID-19). Indian J Pediatr 2020 Apr 13;87(4):281-286. [CrossRef] [Medline]
    5. Li Q, Guan X, Wu P, Wang X, Zhou L, Tong Y, et al. Early transmission dynamics in Wuhan, China, of novel coronavirus-infected pneumonia. N Engl J Med 2020 Mar 26;382(13):1199-1207 [FREE Full text] [CrossRef] [Medline]
    6. Gao Z, Xu Y, Sun C, Wang X, Guo Y, Qiu S, et al. A systematic review of asymptomatic infections with COVID-19. J Microbiol Immunol Infect 2021 Feb;54(1):12-16 [FREE Full text] [CrossRef] [Medline]
    7. Chen N, Zhou M, Dong X, Qu J, Gong F, Han Y, et al. Epidemiological and clinical characteristics of 99 cases of 2019 novel coronavirus pneumonia in Wuhan, China: a descriptive study. Lancet 2020 Feb 15;395(10223):507-513 [FREE Full text] [CrossRef] [Medline]
    8. Wang D, Hu B, Hu C, Zhu F, Liu X, Zhang J, et al. Clinical characteristics of 138 hospitalized patients with 2019 novel coronavirus-infected pneumonia in Wuhan, China. JAMA 2020 Mar 17;323(11):1061-1069 [FREE Full text] [CrossRef] [Medline]
    9. Trouillet-Assant S, Viel S, Gaymard A, Pons S, Richard J, Perret M, COVID HCL Study group. Type I IFN immunoprofiling in COVID-19 patients. J Allergy Clin Immunol 2020 Jul;146(1):206-208.e2 [FREE Full text] [CrossRef] [Medline]
    10. Lan L, Xu D, Ye G, Xia C, Wang S, Li Y, et al. Positive RT-PCR test results in patients recovered from COVID-19. JAMA 2020 Apr 21;323(15):1502-1503 [FREE Full text] [CrossRef] [Medline]
    11. Radpour A, Bahrami-Motlagh H, Taaghi MT, Sedaghat A, Karimi MA, Hekmatnia A, et al. COVID-19 evaluation by low-dose high resolution CT scans protocol. Acad Radiol 2020 Jun;27(6):901 [FREE Full text] [CrossRef] [Medline]
    12. Ciotti M, Maurici M, Pieri M, Andreoni M, Bernardini S. Performance of a rapid antigen test in the diagnosis of SARS-CoV-2 infection. J Med Virol 2021 May 09;93(5):2988-2991 [FREE Full text] [CrossRef] [Medline]
    13. Saberian P, Mireskandari S, Baratloo A, Hasani-Sharamin P, Babaniamansour S, Aliniagerdroudbari E, et al. Antibody rapid test results in emergency medical services personnel during COVID-19 pandemic; a cross sectional study. Arch Acad Emerg Med 2021;9(1):e2 [FREE Full text] [CrossRef] [Medline]
    14. Wang W, Xu Y, Gao R, Lu R, Han K, Wu G, et al. Detection of SARS-CoV-2 in different types of clinical specimens. JAMA 2020 May 12;323(18):1843-1844 [FREE Full text] [CrossRef] [Medline]
    15. Yang J, Zheng Y, Gou X, Pu K, Chen Z, Guo Q, et al. Prevalence of comorbidities and its effects in patients infected with SARS-CoV-2: a systematic review and meta-analysis. Int J Infect Dis 2020 May;94:91-95 [FREE Full text] [CrossRef] [Medline]
    16. Salama C, Han J, Yau L, Reiss WG, Kramer B, Neidhart JD, et al. Tocilizumab in patients hospitalized with COVID-19 pneumonia. N Engl J Med 2021 Jan 07;384(1):20-30. [CrossRef]
    17. Afshar P, Heidarian S, Naderkhani F, Oikonomou A, Plataniotis KN, Mohammadi A. COVID-CAPS: A capsule network-based framework for identification of COVID-19 cases from X-ray images. Pattern Recognit Lett 2020 Oct;138:638-643 [FREE Full text] [CrossRef] [Medline]
    18. Wang L, Lin Z, Wong A. COVID-Net: a tailored deep convolutional neural network design for detection of COVID-19 cases from chest X-ray images. Sci Rep 2020 Nov 11;10(1):19549 [FREE Full text] [CrossRef] [Medline]
    19. Chowdhury MEH, Rahman T, Khandakar A, Mazhar R, Kadir MA, Mahbub ZB, et al. Can AI help in screening viral and COVID-19 pneumonia? IEEE Access 2020;8:132665-132676. [CrossRef]
    20. Zoabi Y, Deri-Rozov S, Shomron N. Machine learning-based prediction of COVID-19 diagnosis based on symptoms. NPJ Digit Med 2021 Jan 04;4(1):3 [FREE Full text] [CrossRef] [Medline]
    21. Hosny KM, Darwish MM, Li K, Salah A. COVID-19 diagnosis from CT scans and chest X-ray images using low-cost Raspberry Pi. PLoS One 2021 May 11;16(5):e0250688 [FREE Full text] [CrossRef] [Medline]
    22. Alorf A. The practicality of deep learning algorithms in COVID-19 detection: application to chest X-ray images. Algorithms 2021 Jun 13;14(6):183. [CrossRef]
    23. Khan AI, Shah JL, Bhat MM. CoroNet: A deep neural network for detection and diagnosis of COVID-19 from chest x-ray images. Comput Methods Programs Biomed 2020 Nov;196:105581 [FREE Full text] [CrossRef] [Medline]
    24. Ucar F, Korkmaz D. COVIDiagnosis-Net: Deep Bayes-SqueezeNet based diagnosis of the coronavirus disease 2019 (COVID-19) from X-ray images. Med Hypotheses 2020 Jul;140:109761 [FREE Full text] [CrossRef] [Medline]
    25. Minaee S, Kafieh R, Sonka M, Yazdani S, Jamalipour Soufi G. Deep-COVID: Predicting COVID-19 from chest X-ray images using deep transfer learning. Med Image Anal 2020 Oct;65:101794 [FREE Full text] [CrossRef] [Medline]
    26. Pham T. Classification of COVID-19 chest X-rays with deep learning: new models or fine tuning? Health Inf Sci Syst 2021 Dec;9(1):2 [FREE Full text] [CrossRef] [Medline]
    27. Kundu R, Das R, Geem ZW, Han G, Sarkar R. Pneumonia detection in chest X-ray images using an ensemble of deep learning models. PLoS One 2021 Sep 7;16(9):e0256630 [FREE Full text] [CrossRef] [Medline]
    28. Maior CBS, Santana JMM, Lins ID, Moura MJC. Convolutional neural network model based on radiological images to support COVID-19 diagnosis: Evaluating database biases. PLoS One 2021 Mar 1;16(3):e0247839 [FREE Full text] [CrossRef] [Medline]
    29. Ismael AM, Şengür A. Deep learning approaches for COVID-19 detection based on chest X-ray images. Expert Syst Appl 2021 Feb;164:114054 [FREE Full text] [CrossRef] [Medline]
    30. Wang D, Mo J, Zhou G, Xu L, Liu Y. An efficient mixture of deep and machine learning models for COVID-19 diagnosis in chest X-ray images. PLoS One 2020 Nov 17;15(11):e0242535 [FREE Full text] [CrossRef] [Medline]
    31. Ahmad F, Farooq A, Ghani MU. Deep ensemble model for classification of novel coronavirus in chest x-ray images. Comput Intell Neurosci 2021 Jan 5;2021:8890226-8890217 [FREE Full text] [CrossRef] [Medline]
    32. MLHC-COVID-19.   URL: http://psuva.com/mlhc [accessed 2023-02-18]
    33. Alqudah AM, Qazan S. Augmented COVID-19 X-ray Images Dataset. Mendeley Data. 2020 Mar 26.   URL: https://data.mendeley.com/datasets/2fxz4px6d8/4 [accessed 2021-11-12]
    34. COVID-19 Database. Italian Society of Medical and Interventional Radiology.   URL: https://sirm.org/covid-19/ [accessed 2021-08-17]
    35. ieee8023 / covid-chestxray-dataset. GitHub. 2021 Mar 22.   URL: https://github.com/ieee8023/covid-chestxray-dataset [accessed 2021-11-12]
    36. Search results for “COVID-19 pneumonia”. Radiopaedia.   URL: https:/​/radiopaedia.​org/​search?lang=us&modality=X-ray&page=1&q=COVID-19+pneumonia&scope=cases&system=Chest [accessed 2021-08-17]
    37. Chest X-Ray Images (Pneumonia). Kaggle.   URL: https://www.kaggle.com/datasets/paultimothymooney/chest-xray-pneumonia [accessed 2021-08-17]
    38. Kanan C, Cottrell GW. Color-to-grayscale: does the method matter in image recognition? PLoS One 2012 Jan 10;7(1):e29740 [FREE Full text] [CrossRef] [Medline]
    39. Masters BR, Gonzalez RC, Woods RE. Book review: Digital Image Processing, third edition. Journal of Biomedical Optics 2009 Mar 01;14(2):029901. [CrossRef]
    40. Singh TM, Singh OM, Singh KM, Sinam T, Singh TR. Image enhancement by adaptive power-law transformations. Bahria University Journal of Information & Communication Technology 2010;3(1):29-37 [FREE Full text]
    41. Chourasiya A, Khare N. A comprehensive review of image enhancement techniques. IJIRG 2019 May 01;8(6):1. [CrossRef]
    42. Chikhalekar AT. Analysis of image processing for digital x-ray. International Research Journal of Engineering and Technology 2016;3(5):1364-1368 [FREE Full text]
    43. Majeeth SS, Babu CNK. Gaussian noise removal in an image using fast guided filter and its method noise thresholding in medical healthcare application. J Med Syst 2019 Jul 12;43(8):280. [CrossRef] [Medline]
    44. Kong NSP, Ibrahim H, Hoo SC. A literature review on histogram equalization and its variations for digital image enhancement. International Journal of Innovation, Management and Technology 2013;4(4):386-389. [CrossRef]
    45. Wongsirichot T, Hanskunatai A. A multi-layer hybrid machine learning model for automatic sleep stage classification. Biomed. Eng. Appl. Basis Commun 2018 Nov 29;30(06):1850041. [CrossRef]
    46. Topîrceanu A, Grosseck G. Decision tree learning used for the classification of student archetypes in online courses. Procedia Computer Science 2017;112:51-60. [CrossRef]
    47. Charbuty B, Abdulazeez AM. Classification based on decision tree algorithm for machine learning. Journal of Applied Science and Technology Trends 2021 Mar 24;2(01):20-28. [CrossRef]
    48. Rizvi S, Rienties B, Khoja SA. The role of demographics in online learning; A decision tree based approach. Computers & Education 2019 Aug;137:32-47. [CrossRef]
    49. Demidova L, Egin M, Tishkin R. A self-tuning multiobjective genetic algorithm with application in the SVM classification. Procedia Computer Science 2019;150:503-510. [CrossRef]
    50. Demidova L, Nikulchev E, Sokolova Y. The SVM classifier based on the modified particle swarm optimization. International Journal of Advanced Computer Science and Applications 2016;7(2):1. [CrossRef]
    51. Shankar K, Lakshmanaprabu SK, Gupta D, Maseleno A, de Albuquerque VHC. Retracted article: optimal feature-based multi-kernel SVM approach for thyroid disease classification. J Supercomput 2018 Jul 02;76(2):1128-1143. [CrossRef]
    52. Alghoul A, Al Ajrami S, Al Jarousha G, Harb G, Abu-Naser SS. Email classification using artificial neural network. International Journal of Academic Engineering Research 2018;2(11):8-14 [FREE Full text]
    53. Maksimenko VA, Kurkin SA, Pitsik EN, Musatov VY, Runnova AE, Efremova TY, et al. Artificial neural network classification of motor-related EEG: an increase in classification accuracy by reducing signal complexity. Complexity 2018:1. [CrossRef]
    54. Kouretas I, Paliouras V. Simplified Hardware Implementation of the Softmax Activation Function. 2019 Presented at: 8th International Conference on Modern Circuits and Systems Technologies (MOCAST); May 13-15, 2019; Thessaloniki, Greece. [CrossRef]
    55. Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research 2014;15:1929-1958 [FREE Full text]
    56. Hossin M, Sulaiman MN. A review on evaluation metrics for data classification evaluations. International Journal of Data Mining & Knowledge Management Process 2015 Mar 31;5(2):01-11. [CrossRef]
    57. García V, Sánchez J, Mollineda R. On the effectiveness of preprocessing methods when dealing with different levels of class imbalance. Knowledge-Based Systems 2012 Feb;25(1):13-21. [CrossRef]
    58. Pedregosa FF, Varoquaux V, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: machine learning in Python. Journal of Machine Learning Research 2011;12:2825-2830 [FREE Full text] [CrossRef]
    59. Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: synthetic minority over-sampling technique. Journal of Artificial Intelligence Research 2002 Jun 01;16:321-357. [CrossRef]
    60. Pan T, Zhao J, Wu W, Yang J. Learning imbalanced datasets based on SMOTE and Gaussian distribution. Information Sciences 2020 Feb;512:1214-1233. [CrossRef]
    61. tanthanakorn2541 / MLHC-COVID-19. GitHub.   URL: https://github.com/tanthanakorn2541/MLHC-COVID-19/tree/master/Dataset_Original [accessed 2023-02-19]

    ‎
    AUC: area under the curve
    CNN: convolutional neural network
    CT: computed tomography
    CXR: chest x-ray
    DT: decision tree
    FN: false negative
    FP: false positive
    MLHC: multilayer hybrid classification model
    NN: neural network
    ROC: receiver operator characteristic
    RT-PCR: reverse transcription polymerase chain reaction
    SARS: severe acute respiratory syndrome
    SMOTE: synthetic minority oversampling technique
    SVM: support vector machine
    TN: true negative
    TP: true positive
    WHO: World Health Organization

    Edited by A Mavragani; submitted 31.08.22; peer-reviewed by M Kapsetaki; comments to author 30.01.23; revised version received 10.02.23; accepted 13.02.23; published 28.02.23

    Copyright

    ©Thanakorn Phumkuea, Thakerng Wongsirichot, Kasikrit Damkliang, Asma Navasakulpong. Originally published in JMIR Formative Research (https://formative.jmir.org), 28.02.2023.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Formative Research, is properly cited. The complete bibliographic information, a link to the original publication on https://formative.jmir.org, as well as this copyright and license information must be included.