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.

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.

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.

where, if G_intensity(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.

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.

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].

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.

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.

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.

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.

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.