Identification of COVID-19 Chest X- Ray Images using a Multi-layer Hybrid Classification Model

doi:10.21203/rs.3.rs-1512775/v1

Download PDF

Research Article

Identification of COVID-19 Chest X- Ray Images using a Multi-layer Hybrid Classification Model

https://doi.org/10.21203/rs.3.rs-1512775/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background: COVID-19 diagnoses during the pandemic require fast and accurate identification of chest X-Ray (CXR) images. Image processing techniques are critical in accurately diagnosing diseases. This study establishes a method for image analysis and classification using the Multi-Layer Hybrid Classification (MLHC) technique on COVID-19, non-COVID-19 (viral and bacterial pneumonia), and healthy CXR images.

Methods: In this retrospective study, five public datasets of 4,200 CXR images were analyzed using several machine learning techniques which include Decision Tree (DT), Support Vector Machine (SVM) and Neural Networks (NN). Pre-processing was performed by image segmentation, image enhancement, and feature extraction techniques. The best performing classifiers were identified based on evaluation criterion and ensembled into MLHC for COVID-19 (MLHC-COVID-19).

Results: MLHC-COVID-19 was trained using ten-fold cross-validation. For the unseen test, MLHC-COVID-19 showed promising performance achieving an accuracy, sensitivity, and specificity of 96.2%, 96.2%, and 98.1%, respectively. Therefore, MLHC-COVID-19 can effectively classify CXR images enhancing the accuracy of CXR image interpretation.

Conclusion: Rigorous comparative studies used more than ten advanced convolutional neural network (CNN) based classification models. The results showed that the proposed method outperforms the comparative models in classification. Therefore, MLHC-COVID-19 can be a potential practical computer-aided diagnostic (CADx) tool for COVID-19 CXR classification. Additionally, a web-based MLHC-COVID-19 CADx is available at http://psuva.com/mlhc.

COVID19

CXR

Machine Learning

Hybrid Model

Artificial Intelligence

Viruses, bacteria, and fungi can all cause lung infection or pneumonia [1]. Chest X-Ray (CXR) examination is considered gold standard in diagnosing pneumonia. With the recent widespread of SARS-CoV-2, differentiation between COVID-19 pneumonia and non-COVID-19 pneumonia (other viruses or bacteria) is a challenging task. In resource-constrained situations, computer-aided diagnosis (CADx) may help clinicians identify COVID-19 pneumonia using CXR images [2] [3]. With advancements in computer vision and machine learning, CADx of medical images are becoming increasingly viable. CXR image analysis using state-of-the-art machine learning approaches has shown promising results in identifying patients with pneumonia related to COVID-19 infection [4–6]. The researchers in [7] proposed COVID-CAPS, based on capsule networks, as a modeling framework. COVID-CAPS can perform effectively on small datasets. The dataset was retrieved from the NIH repository and consisted of 94,323 CXR images of common thoracic diseases. The reported method achieved an accuracy, sensitivity, and specificity, respectively, of 95.7%, 90.0%, and 95.8%; however, the authors did not report the details of the dataset used. CoroNet model [8] was proposed to classify COVID-19 CXR images. Accuracy of 89.6% and 95.0% for classification, respectively of four and three classes, was achieved with a dataset consisting of 284 COVID-19, 327 viral pneumonia, 330 bacterial pneumonia, and 310 normal X-ray images. A similar work in [9] implemented VGG16 to distinguish COVID-19 infections from non-infected and healthy cases. They developed two machine learning models using 6,523 CXR images (250 COVID-19 CXR images). The first model obtained an accuracy of 96.0% when discriminating between healthy and generic pulmonary diseases, whereas the second model achieved an accuracy of 98.0% for distinguishing between generic pulmonary diseases and COVID-19. In [10], an image augmentation technique was used for COVID-19 CXR images. They fine- tuned a pre-trained model, SqueezeNet, using Bayesian optimization. Their proposed model classified COVID-19 on CXR images. The study consisted of 4,290 pneumonia, 1,583 normal, and 76 COVID-19 CXR images. Furthermore, pneumonia and normal classes were partially split to balance the dataset. The authors obtained an accuracy and an F-measure of 98.3% and 98.3%, respectively, during testing. Whereas [11] employed a transfer learning approach, image augmentation was utilized on 423 COVID-19, 1,485 viral pneumonia, and 1,579 normal CXR images. The authors proposed a different classification model. Two different classifications — normal and COVID-19 (two classes) and normal, viral pneumonia, and COVID-19 (three classes)–were identified. The two classes classifications respectively achieved an accuracy, sensitivity, and F-measure of 99.7%, 99.7%, and 99.7%. Moreover, the three classes classifications respectively obtained an accuracy, sensitivity, and F-measure of 97.4%, 97.4%, and 97.4%. A transfer learning approach for identifying the COVID-19 disease in CXR images [12] was presented using a COVID-19 dataset divided into 2,084 training and 3,100 test images. The study performed four different transfer learning techniques: ResNet18, ResNet50, SqueezeNet, and DenseNet-121. SqueezeNet achieved the highest specificity (92.9%) and sensitivity (98.0%); however, the accuracy rates of different transfer learning techniques were not reported. In [13], the researchers proposed an algorithm to iteratively prune the fine-tuning of the DL model ensembles. This algorithm was used to determine the optimal number of neurons in the convolutional layers. The dataset consisted of 7,595 normal images, 3,012 pneumonia images (type unknown), 2,780 bacterial pneumonia images, and 313 COVID-19 images. The authors achieved the highest accuracy of 97.4% for a three-class classification task.

Researchers in [14] used a generative adversarial network (GAN) [15] to increase the number of CXR images and to prevent overfitting because the dataset contained 198 COVID-19 and 210 healthy images. They performed a feature selection technique using principal component analysis (PCA) in combination with convolutional neural networks (CNNs). The proposed method achieved a high accuracy of 1.0 showing promising results for small datasets. This study should be re-conducted using a larger dataset. The authors in [16] developed the VGGNet model using a histogram-oriented gradient (HOG) in combination with a CNN to detect COVID-19. The CNN classifiers were trained using five-fold cross-validation to evaluate the model’s performance. This study examined 3,111 normal and 1,979 COVID-19 CXR images. The fusion model attained an accuracy of 99.5%. The authors in [17] used the optimized genetic algorithm-extreme learning machine (OGA-ELM) approach from [18] for COVID-19 detection on CXR images. They used HOG and PCA to extract the features of 94 COVID-19-infected images and 94 uninfected images. The evaluation result showed accuracy of 1.00 in classifying two classes with reduced computation time. However, a small dataset was employed, which limited the scope of their research.

In [19], deep feature extraction and fine-tuning of a pre-trained CNN, as well as an end-to-end trained new CNN model, were proposed as three deep learning approaches for COVID-19 detection. The researchers experimented on a CXR dataset containing 180 COVID-19 and 200 healthy images. The deep features extracted from the ResNet50 model and support vector machine (SVM) classifier outperformed other approaches with an accuracy of 94.7%. A similar work in [20] adopted five different pre-trained deep models that classified using various ma- chine learning classifiers (i.e., decision tree, random forest, AdaBoost, bagging, and SVM). They used a dataset of 1102 CXR images, consisting of 565 normal and 537 COVID-19 images for their experiments. The dataset was divided into training and test sets in a ratio of 70:30. The results showed that the Xception combined with the SVM classifier was the most efficient, with accuracy, sensitivity, and specificity of 99.3%, 99.2%, and 99.3%, respectively. Furthermore, in [21] the authors proposed a hybrid ensemble model for identifying COVID-19 infections in CXRs. The proposed method was developed by combining MobileNet and InceptionV3. Both deep models were fine-tuned for the learned weights. The authors performed a four-fold cross-validation on the dataset of 1,050 normal, 1,050 viral pneumonia, 1,050 bacterial pneumonia, and 1,050 COVID-19 CXR images. The test set result showed an accuracy, precision, and specificity of 94.2%, 89.9%, and 88.3%, respectively. Table 1 shows a summary of related studies on COVID-19 CXR images.

Table 1

Related studies of the COVID-19 CXR Images.
Method	Dataset (C = COVID-19 : B = Bacteria Pneumonia : V = Virus Pneumonia : H = Healthy : P = Pneumonia)	No. of Classified Classes	Cross Validation	Ref
CNN-based Capsule Networks	Not specified	5	No	[7]
CoroNet (Xception)	C = 284 : B = 330 : V = 327 : H = 310	3 and 4	4-fold	[8]
VGG-16	C = 250 : H = 3,520 : P = 2,753	2	No	[9]
Bayes-SqueezeNet	C = 76 : H = 1,583 : P = 4,290	3	No	[10]
Different pre-trained CNN model	C = 423 : V = 1,485 : H = 1,579	2 and 3	5-fold	[11]
Different pre-trained CNN model	C = 184 : H = 5,000	2	5-fold	[12]
Iteratively pruned deep learning model ensembles	C = 313 : B = 2,780 : H = 7,595 : P = 3,012	3	No	[13]
PCA + CNN	C = 198 : H = 210	2	No	[14]
Fusion features + VGG19 pre-trained model	C = 1,979 : H = 3,111	2	5-fold	[16]
HOG + PCA + OGA-ELM	C = 94 : H = 94	2	No	[17]
ResNet50 for Features extraction + SVM	C = 180 : H = 200	2	No	[19]
Xception + SVM	C = 537 : H = 565	2	10-fold	[20]
Hybrid ensemble model	C = 1,050 : V = 1,050 : B = 1,050 : H = 1,050	4	4-fold	[21]
MLHC-COVID-19	C = 1,050 : non-COVID-19 = 2100 : H = 1,050	3	10-fold	Our method

On literature review, most of the reported studies used DL and a pre-trained model with data augmentation to detect COVID-19 from CXR images. However, to the best of our knowledge, only a few studies have used multi-layer classifiers for detecting COVID-19 based on CXR images. Also, there is a lack of traditional machine learning models conducted on open source with x-ray images. The MLHC-COVID-19 model aims to address these shortcomings.

The general overview of the proposed MLHC-COVID-19 for identifying CXR images of COVID-19 infection is shown in Fig. 1. The block diagram depicts the vital processes embedded in the model. The following subsections will discuss each process.

COVID-19 CXR images dataset

In this retrospective study, five datasets of CXR images were used and made publicly available by Mendeley [22], the Italian Society of Medical and Interventional Radiology (SIRM) [23], Github [24], Radiopaedia [25], and Kaggle [26]. The number of images used in this study were 4,200 divided in three classes: (a) healthy, (b) non-COVID-19 (viral pneumonia + bacterial pneumonia), and (c) COVID-19. Each image was annotated by medical specialists. The first four sources contained 1,050 CXR images of COVID-19 infected patients (912 + 138). The last source contained 3,150 CXR images of healthy, viral pneumonia, and bacterial pneumonia patients (1,050 healthy, 2,100 non-COVID-19). Table 2 lists the CXR image datasets used in this study and the samples of the CXR images in the combined dataset are shown in Fig. 2.

Table 2

CXR images datasets
Open Source	Classes	Number of Images
Kaggle [26]	Healthy	1050
Kaggle [26]	non-COVID-19 (Viral + Bacterial)	2100 (1050 + 1050)
Mendeley dataset [22]	COVID-19	912
SIRM [23], Github [50] and Radiopaedia [25]	COVID-19	138

Image Preprocessing

In this study, image pre-processing consists of three steps: image segmentation, image analysis, and feature extraction as shown in Fig. 3. For image segmentation, we designed three sub-steps: removal of the partial black background, image resizing, and grayscale conversion. The combined dataset had different formats and a partial black background in the CXR images. This was caused by many factors such as the filming position, types of X-ray machines used, etc. The black background affected the classification performance therefore it was removed. In Fig. 4, an example of a black background completely removed. Furthermore, the CXR images were of different dimensions; consequently, all the images were resized to an identical size of 500 × 500 pixels. The last image segmentation step is converting the images to grayscale to reduce the image features. Reducing the image features improves the classification result and mimics the complexity of the algorithm [27]. The grayscale formula is given by Eq. 1.

$${G}_{intensity}\left(x,y\right)= 0.2989\text{*}f\left(x,y,R\right)+0.5871\text{*}f\left(x,y,G\right)+0.1140\text{*}f\left(x,y,B\right)$$

where ${G}_{intensity}$ (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 [28].

The second image preprocessing step 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 Eq. 2 [29, 30]. In addition, the two-dimensional Gaussian filter technique was selected to reduce the Gaussian and salt-and-pepper noise [31]. The Gaussian filter technique is given by Eq. 3. In Fig. 5, the CXR image after image enhancement.

$$s=c*{r}^{\gamma }$$

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.

$${G}_{Filter}\left(x,y\right)= \frac{1}{2\pi {\partial }^{2}}{e}^{-\frac{{x}^{2}+{y}^{2}}{2{\partial }^{2}}}$$

where ∂² is the variance of the Gaussian filter with a 3 x 3 kernel size, x, and y are the horizontal and vertical axes of the kernel size [32].

The last image preprocessing step is feature extraction for which histogram analysis and L2-normalization were performed. Histogram analysis reduces the image features by retrieving vital image statistics. Specifically, low-intensity values produce a dark-toned image and vice versa [33] (see Algorithm 2 in Appendix A). After the histogram analysis, each feature in the histogram had different scales. Subsequently, to obtain a standard scale, we performed L2-Normalization. L2-Normalization or the Euclidean norm normalizes the features of the histogram to the same scale using Eq. 4. In Fig. 6, the output of feature extraction.

$$L2= \frac{X}{\sqrt{{X}_{i}^{2} + {X}_{i+1}^{2} + {X}_{i+2}^{2} + \dots + {X}_{i+n}^{2}}}$$

Where X is a feature of the histogram.

Multi-layer Hybrid Classification Model

First, the original MLHC was developed to classify the sleep stages automatically by our research team. Each layer in the MLHC is a binary classification model that uses different machine-learning techniques [34]. The MLHC-COVID-19 method in this study was designed with two layers. The first differentiated unhealthy or infected subjects (viral, bacterial, and COVID-19 infection) from healthy subjects. The second classified COVID-19 and non-COVID-19 (viral and bacterial) subjects. Figures 7 shows pseudocode of the MLHC-COVID-19.

Three machine-learning techniques, DT, SVM, and NN were experimentally evaluated as candidates for embedding into MLHC-COVID19. DT is a well-known supervised machine-learning method. Each node represents a condition for deciding on data classification, where various branches of trees represent the results from testing and the leaves of the DT represent the classification [35, 36]. DT is one of the simplest techniques to understand and is suitable for classification tasks [37]. SVM is a supervised machine-learning method with promising performance in statistical classifications [38]. It distinguishes data by finding hyperplanes as separators. Identifying hyperplanes iterates toward the best line during the training [39]. The radial basis function (RBF) was selected as the kernel function [40]. Finally, NN are types of mathematical models for processing data with connected computation nodes that mimic the functions of biological neural networks [41]. They build complex models between the inputs and outputs with high efficiency [42]. We designed four fully connected layers (dense layers) and two 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 [43]. In addition, dropout layers with an 0.2 ratio were interpolated between the dense layers [44]. Figures 8 shows the MLHC-COVID-19 flowchart.

Performance Evaluation

The training and testing datasets were divided using an 80:20 ratio. The training dataset was used to train the model, which was evaluated using ten-fold cross- validation. The testing or unseen dataset was used to select the best model. The performance measures were accuracy, sensitivity, specificity, precision, F-measure, and the area under the curve (AUC) [45, 46]. The performance measure formulae are given in Eq. 5–10., and the measured performances include true positive (TP), true negative (TN), false positive (FP), and false negative (FN). In addition, the classification performance was evaluated using the area of the receiver operating characteristic curve (AUC of ROC). Python module 3.8.3 known as Scikit-learn was used for the machine learning algorithms [47] with an NVIDIA GeForce-960 M GPU, 4 GB GDDR5, on-board memory with an Intel Core i7-6700 HQ (2.60 GHz, 6 MB L3 Cache, approximately 3.50 GHz), 8 GB DDR4 RAM, and a 1 TB hard drive.

$$Accuracy = \frac{TP+TN}{\left(TP+FP\right)+\left(TN+FN\right)}$$

$$Sensitivity = \frac{TP}{\left(TP+FN\right)}$$

$$Specificity = \frac{TN}{(TN+FP)}$$

$$Precision = \frac{TP}{(TP+FP)}$$

$$F Measure = \frac{2TP}{(2TP+FP+FN)}$$

$AUC = \frac{1}{2}(\frac{TP}{TP+FN}+ \frac{TN}{TN+FP})$ (10

The MLHC-COVID-19 design left the distribution of data in each layer imbalanced, potentially causing bias in the evaluation. The synthetic minority oversampling technique (SMOTE) was applied to balance the classes [48]. The SMOTE synthesized new data from the existing data using k-nearest neighbors (kNN) and inserted them into the original dataset [49]. This technique alleviated the imbalanced class problem in the MLA. Table 3 shows the number of images of the original and applied-SMOTE datasets in each class based on the MLHC-COVID-19 design. For example, there were 840 subjects in the Healthy class and 2,520 subjects in the Non- healthy class. The Healthy class became a minority class in Layer I. We synthesized 1,680 healthy instances using the SMOTE to balance both classes.

Table 3

Comparison of the number of instances of the original and applied-SMOTE datasets.
MLHC-COVID-19	Class	No. of Images in Training set (80%)		No. of Images in Unseen Testing set (20%)
MLHC-COVID-19	Class	Original	SMOTE	Original	SMOTE
Layer I	Healthy	840	2520	210	630
Layer I	Unhealthy	2520	2520	630	630
Layer II	COVID-19	840	1680	210	420
Layer II	Non-COVID-19	1680	1680	420	420

Three machine-learning techniques, DT, SVM, and NN were investigated. The best classification model for each layer was selected for the MLHC-COVID-19. Table 4 shows the classification results for the training and testing. During the training, ten-fold cross-validation was performed. First, the NN achieved the highest accuracy at 98.8% during the training, followed by the SVM and DT in Layer I. In addition, the NN achieved the highest position with a sensitivity of 99.2%, specificity of 98.5%, precision of 98.5%, F measure of 98.9%, and AUC of 99.5%. 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 (AUC of 98.5%). The second best was the NN followed by the DT. Subsequently, Fig. 9 (a) and (b) graphically represent the averaged zero-fold cross-validation’s ROC after training. The NN achieved the best classification result in Layer I and the SVM outperformed all other models in Layer II. Second, as previously mentioned, 20% of the dataset was used for testing. The classification results in the unseen test also confirm that the NN and SVM are the best classifiers in Layers I and II, respectively. Table 4 shows the highest accuracy of 96.7% (NN) in Layer I and 98.7% (SVM) in Layer II.

Table 4

Classification results.
Layer	Performance Metric	Train (10-fold cross-validation)			Unseen Test (Hold-out)
Layer	Performance Metric	DT	SVM	NNs	DT	SVM	NNs
Layer I (Healthy, Unhealthy)	Accuracy (%)	86.3	96.2	98.8*	83.6	94.2	96.7*
	Sensitivity (%)	71.9	97.2	99.2*	75.4	92.4	95.7*
	Specificity (%)	91.2	95.3	98.5*	91.7	96.0	97.6*
	Precision (%)	72.8	95.4	98.5*	90.1	95.7	97.6*
	F Measure (%)	72.3	96.3	98.9*	82.1	94.1	96.6*
	AUC (%)	81.9	96.2	99.5*	83.6	94.2	96.7*
	Time (sec)	1.108*	1.201	24.888	< 1.000	< 1.000	< 1.000
Layer II (COVID-19, Non-COVID-19)	Accuracy (%)	93.6	98.5*	97.9	94.6	98.7*	98.5
	Sensitivity (%)	95.0	98.1*	97.7	94.8	98.3*	98.1
	Specificity (%)	90.8	98.7*	98.1	94.5	99.0*	98.8
	Precision (%)	95.4	98.7*	98.1	94.5	99.0*	98.8
	F Measure (%)	95.2	98.4*	97.9	94.6	98.7*	98.4
	AUC (%)	92.9	98.4*	97.9	94.6	98.7*	98.5
	Time (sec)	0.749	0.275*	18.183	< 1.000	< 1.000	< 1.000
* Best classification results.

For the MLHC-COVID-19 method, NN were applied in the first layer and the SVM in the second. In Fig. 10, the complete confusion matrix of the MLHC- COVID-19 in the unseen test. It correctly classified 415 COVID-19 cases. Two COVID-19 cases were classified as non-COVID-19 (viral and bacterial pneumonia) and three COVID-19 cases were incorrectly classified. Conversely, eight cases (seven non-COVID-19 and one healthy) were incorrectly classified as COVID-19 cases. The MLCM-COVID-19 classification results for the unseen test show comprehensively superior performance over the other techniques. This is discussed below and considers the accuracy, sensitivity, specificity, and AUC of 96.2%, 96.2%, 98.1%, and 97.1%.

Deep Learning vs Hybrid Machine Learning Technique

COVID-19 can be diagnosed using CXR images. Many studies have reported various pre-processing, feature extraction, and classification methods, among which customized CNN and ensemble learning are the foremost deep learning approaches. These approaches diagnose efficiently; however, are difficult to implement practically owing to their complex processing. Nevertheless, the MLHC-COVID-19 can be used in CADx especially with large datasets because it is simpler and has faster processing time than its counterparts. The results shown below were slightly different using the same resources.

Misdiagnoses can expose patients to risks. This study's objective was to develop a model with perfect or near-perfect accuracy levels. Compared to relevant studies, it achieved an accuracy of over 90%, which is considered very high from a statistical perspective. The results in Table 5 compared these performances using the same resources and showed that the proposed model outperformed the hybrid ensemble model. In [21], researchers proposed a hybrid ensemble model for multi-class classification, with its evaluation result on the test set showing an accuracy, sensitivity, precision, and F-measure of 94.2%, 88.4%, 89.9%, and 88.6%, respectively. On the same dataset, the MLHC-COVID-19 produced classification with an accuracy of 96.2%, an AUC value of 97.1%, and specificity of 98.1%.

Table 5

Comparison of the COVID-19 classification results.
No. of Classes	Accuracy (%)	Sensitivity (%)		Precision (%)	F Measure (%)	AUC (%)	Author
5	95.7	90.0	95.8	-	-	97.0	[7]
3	95.0	96.9	97.5	95.0	95.6	-	[8]
3	98.3	-	99.1	-	98.3	-	[10]
3	97.9	97.9	98.8	98.0	97.9	-	[11]
2	-	98.0	90.0	-	-	-	[12]
3	99.0	99.0	-	99.0	99.0	99.7	[13]
2	100	100	-	100	100	-	[14]
2	99.5	93.7	95.7	-	-	-	[16]
2	100	-	100.0	100	100	-	[17]
2	94.7	91.1	98.0	97.6	94.3	-	[19]
2	99.3	99.0	99.4	99.3	99.3	99.3	[20]
4	94.2	88.4	-	89.9	88.6	-	[21]
3	96.2	96.2	98.1	96.2	96.2	97.1	MLHC-COVID-19

Currently, the COVID-19 pandemic crisis is ongoing owing to the rapid spread of viral infection. Therefore, a rapid COVID-19 diagnosis is essential. This study has developed MLHC-COVID-19, a hybrid classification model, to detect lung infections automatically. It incorporates three ML techniques: DT, SVM, and NN, which were trained, validated, and tested for classifying no infection, pneumonia, and COVID- 19 patients using CXR images. Furthermore, the dataset was divided according to a newly proposed multi-layer architecture. Considering the first layer, the NN slightly outperformed the other techniques, while in second, the SVM outperformed the other two techniques.

The first layer used as NN and the second as SVM equipped with a newly proposed algorithm produced the above results. The classification accuracy, sensitivity, specificity, and precision on healthy, viral and bacterial pneumonia, and COVID-19 images were 96.2%, 96.2%, 98.1%, and 96.2%, respectively. The MLHC- COVID-19 classification results were acceptable compared to other techniques.

Web-based Computer-aided Diagnosis Tool for COVID-19 from CXR Images

As mentioned above, the MLHC-COVID-19 is a simple to apply computer-aided diagnosis tool, which is highly accurate. Considering the pandemic, this CADx tool must be easily accessible; therefore, a web-based solution is available. The prototype web-based CADx brings the MLHC-COVID-19 technique into the clinical setting as a useful tool for radiologists to improve COVID-19 diagnostic accuracy. The tool allows users to upload CXR images, which are then diagnosed to evaluate for non-infection, pneumonia, or COVID-19. The system is available at http://psuva.com/mlhc.

Limitation and Future work

The limitation in the proposed study is that raw images must be pre-processed before being fed into a machine learning model otherwise the result may be ineffective. Image pre-processing that removes noisy or distorted pixels is critical for obtaining meaningful information and accurate classification. The classification model diagnosed infections from the appropriately preprocessed CXR images efficiently. Interestingly, histogram analysis of CXR images and L2-normalization consumed little time in training the data because it does not localize the regions where the CXR images show abnormalities. The classification model has a limitation that could be improved in future work. Histogram analysis was used to obtain complete images. Other techniques classify COVID-19 very well, however a few studies have used multi-layer classifiers to detect COVID-19 in chest X-ray images. The proposed techniques uses these resources without requiring extended processing. In future, plans are to use image data augmentation techniques on CXR images to improve the classification accuracy. Moreover, the study is limited because it is only classified into three classes. Finally, we plan to expand the dataset and develop a model that can classify into four classes by adding a new layer for classifying viral and bacterial pneumonia to the multi-layer architecture.

MLHC

Multi-Layer Hybrid Classification.

CXR

Chest X-ray.

CNN

Convolutional neural networks.

Deep Learning.

CADx

Computer-aided diagnosis.

Decision Tree.

SVM

Support Vector Machine.

Neural Networks.

AUC

Area Under the Curve.

Acknowledgements

Not applicable.

Authors’ contributions

Conceptualization, Thanakorn Poomkur and Thakerng Wongsirichot., methodology, Thanakorn Poomkur and Thakerng Wongsirichot., software, Thanakorn Poomkur., validation, Thanakorn Poomkur and Thakerng Wongsirichot., formal analysis, Thanakorn Poomkur and Thakerng Wongsirichot., writing—original draft preparation, Thanakorn Poomkur and Thakerng Wongsirichot., writing—review and editing, Thanakorn Poomkur and Thakerng Wongsirichot., supervision, Thakerng Wongsirichot., funding acquisition, Thanakorn Poomkur. All authors have read and agreed to the published version of the manuscript.

Funding

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 no. 2046735, Ministry of Higher Education, Science, Research and Innovation, Thailand.

Availability of data and materials

The datasets used for our study can be found in the following repository: https://github.com/tanthanakorn2541/MLHC-COVID-19. All code for image processing and machine learning models used in this study is also included in the above repository.

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Author details

1 College of Digital Science, Prince of Songkla University, Songkhla, Thailand. 2Division of Computational Science, Faculty of Science, Prince of Songkla University, Songkhla, Thailand.

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;384.
Hwang EJ, Kim KB, Kim JY, Lim JK, Nam JG, Choi H, et al. COVID-19 pneumonia on chest X-rays: Performance of a deep learning-based computer-aided detection system. PLoS One. 2021;16 6 June.
Pereira RM, Bertolini D, Teixeira LO, Silla CN, Costa YMG. COVID-19 identification in chest X-ray images on flat and hierarchical classification scenarios. Comput Methods Programs Biomed. 2020;194.
Kassania SH, Kassanib PH, Wesolowskic MJ, Schneidera KA, Detersa R. Automatic Detection of Coronavirus Disease (COVID-19) in X-ray and CT Images: A Machine Learning Based Approach. Biocybern Biomed Eng. 2021;41.
Shalev-Shwartz S, Ben-David S. Understanding machine learning: From theory to algorithms. 2013.
Yousri D, Abd Elaziz M, Abualigah L, Oliva D, Al-qaness MAA, Ewees AA. COVID-19 X-ray images classification based on enhanced fractional-order cuckoo search optimizer using heavy-tailed distributions. Appl Soft Comput. 2021;101.
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;138:638–43.
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;196.
Brunese L, Mercaldo F, Reginelli A, Santone A. Explainable Deep Learning for Pulmonary Disease and Coronavirus COVID-19 Detection from X-rays. Comput Methods Programs Biomed. 2020;196.
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;140.
Chowdhury MEH, Rahman T, Khandakar A, Mazhar R, Kadir MA, Mahbub Z Bin, et al. Can AI Help in Screening Viral and COVID-19 Pneumonia? IEEE Access. 2020;8:132665–76.
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;65.
Rajaraman S, Siegelman J, Alderson PO, Folio LS, Folio LR, Antani SK. Iteratively Pruned Deep Learning Ensembles for COVID-19 Detection in Chest X-Rays. IEEE Access. 2020;8:115041–50.
Rasheed J, Hameed AA, Djeddi C, Jamil A, Al-Turjman F. A machine learning-based framework for diagnosis of COVID-19 from chest X-ray images. Interdiscip Sci Comput Life Sci. 2021;13:103–17.
Goodfellow IJ, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, et al. Generative adversarial nets. In: Advances in Neural Information Processing Systems. 2014.
Nur-a-alam, Ahsan M, Based MA, Haider J, Kowalski M. COVID-19 detection from chest X-ray images using feature fusion and deep learning. Sensors. 2021;21.
Albadr MAA, Tiun S, Ayob M, Al-Dhief FT, Omar K, Hamzah FA. Optimised genetic algorithm-extreme learning machine approach for automatic COVID-19 detection. PLoS One. 2020;15 12 December.
Albadr MAA, Tiun S, Ayob M, AL-Dhief FT. Spoken language identification based on optimised genetic algorithm–extreme learning machine approach. Int J Speech Technol. 2019;22.
Ismael AM, Şengür A. Deep learning approaches for COVID-19 detection based on chest X-ray images. Expert Syst Appl. 2021;164.
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;15 11 November:1–15.
Ahmad F, Farooq A, Ghani MU. Deep Ensemble Model for Classification of Novel Coronavirus in Chest X-Ray Images. Comput Intell Neurosci. 2021;2021.
Alqudah AM, Qazan S. Augmented COVID-19 X-ray Images Dataset - Mendeley Data. 2020. https://data.mendeley.com/datasets/2fxz4px6d8/4. Accessed 12 Nov 2021.
Italian Society of Medical and Interventional Radiology. Covid-19 - SIRM. https://sirm.org/covid-19/. Accessed 17 Aug 2021.
Cohen JP, Morrison P, Dao L, Roth K, Duong TQ, Ghassemi M. COVID-19 Image Data Collection: Prospective Predictions Are the Future. J Mach Learn Biomed Imaging. 2020;2020:2–3.
Radiopaedia. COVID-19 pneumonia | Search | Radiopaedia.org. https://radiopaedia.org/search?lang=us&modality=X-ray&page=1&q=COVID-19+pneumonia&scope=cases&system=Chest. Accessed 17 Aug 2021.
Kaggle. Chest X-Ray Images (Pneumonia) | Kaggle. kaggle. 2020. https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia. Accessed 17 Aug 2021.
Kanan C, Cottrell GW. Color-to-grayscale: Does the method matter in image recognition? PLoS One. 2012;7.
Gonzalez RC, Woods RE, Masters BR. Digital Image Processing, Third Edition. J Biomed Opt. 2009;14.
T.Romen Singh1, O.Imocha Singh1, Kh. Manglem Singh2 TS and TRS. Image Enhancement by Adaptive Power-Law Transformations. Bahria Univ J Inf Commun Technol. 2010;3:29–37.
Chourasiya A, Khare N. A Comprehensive Review Of Image Enhancement Techniques. Int J Innov Res Growth. 2019;8.
Chikhalekar AT. Analysis of Image Processing for Digital X-Ray. Int Res J Eng Technol. 2016;3.
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;43.
Kong N. A Literature Review on Histogram Equalization and Its Variations for Digital Image Enhancement. Int J Innov Manag Technol. 2013. https://doi.org/10.7763/ijimt.2013.v4.426.
Wongsirichot T, Hanskunatai A. A MULTI-LAYER HYBRID MACHINE LEARNING MODEL for AUTOMATIC SLEEP STAGE CLASSIFICATION. Biomed Eng - Appl Basis Commun. 2018;30.
Topîrceanu A, Grosseck G. Decision tree learning used for the classification of student archetypes in online courses. In: Procedia Computer Science. 2017.
Charbuty B, Abdulazeez A. Classification Based on Decision Tree Algorithm for Machine Learning. J Appl Sci Technol Trends. 2021;2.
Rizvi S, Rienties B, Khoja SA. The role of demographics in online learning; A decision tree based approach. Comput Educ. 2019;137.
Demidova LA, Egin MM, Tishkin R V. A self-tuning multiobjective genetic algorithm with application in the SVM classification. In: Procedia Computer Science. 2019.
Demidova L, Nikulchev E, Sokolova Y. The SVM Classifier Based on the Modified Particle Swarm Optimization. Int J Adv Comput Sci Appl. 2016;7.
Shankar K, Lakshmanaprabu SK, Gupta D, Maseleno A, de Albuquerque VHC. Optimal feature-based multi-kernel SVM approach for thyroid disease classification. J Supercomput. 2020;76.
Alghoul A, Ajrami S Al, Jarousha G Al, Harb G, Abu-Naser SS. Email Classification Using Artificial Neural Network. Int J Acad Eng Res. 2018;2:8–14.
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;2018.
Kouretas I, Paliouras V. Simplified Hardware Implementation of the Softmax Activation Function. 2019 8th Int Conf Mod Circuits Syst Technol MOCAST 2019. 2019. https://doi.org/10.1109/MOCAST.2019.8741677.
Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: A simple way to prevent neural networks from overfitting. J Mach Learn Res. 2014;15.
M H, M.N S. A Review on Evaluation Metrics for Data Classification Evaluations. Int J Data Min Knowl Manag Process. 2015;5.
García V, Sánchez JS, Mollineda RA. On the effectiveness of preprocessing methods when dealing with different levels of class imbalance. Knowledge-Based Syst. 2012;25.
Pedregosa FABIANPEDREGOSA F, Michel V, Grisel OLIVIERGRISEL O, Blondel M, Prettenhofer P, Weiss R, et al. Scikit-learn: Machine Learning in Python Gaël Varoquaux Bertrand Thirion Vincent Dubourg Alexandre Passos PEDREGOSA, VAROQUAUX, GRAMFORT ET AL. Matthieu Perrot. J Mach Learn Res. 2011;12:2825–30.
Chawla N V., Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: Synthetic minority over-sampling technique. J Artif Intell Res. 2002;16.
Pan T, Zhao J, Wu W, Yang J. Learning imbalanced datasets based on SMOTE and Gaussian distribution. Inf Sci (Ny). 2020;512.
Cohen JP, Morrison P, Dao L, Roth K, Duong TQ, Ghassemi M. ieee8023/covid-chestxray-dataset: We are building an open database of COVID-19 cases with chest X-ray or CT images. https://github.com/ieee8023/covid-chestxray-dataset. Accessed 12 Nov 2021.

Appendix A is not available with this version

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Identification of COVID-19 Chest X- Ray Images using a Multi-layer Hybrid Classification Model

Status:

Version 1

Abstract

Figures

Introduction

Methodology

Experimental Results

Results And Discussion

Conclusion

Abbreviations

Declarations

References

Supplementary

Additional Declarations

Status:

Version 1