A Novel Method for Multivariant Pneumonia Classification based on Hybrid CNN-PCA Based Feature Extraction using Extreme Learning Machine with Chest X-Ray Images

In this era of COVID19, proper diagnosis and treatment for pneumonia are very important. Chest X-Ray (CXR) image analysis plays a vital role in the reliable diagnosis of pneumonia. An experienced radiologist is required for this. However, even for an experienced radiographer, it is quite difficult and time-consuming to diagnose due to the fuzziness of CXR images. Also, identification can be erroneous due to the involvement of human judgment. Hence, an authentic and automated system can play an important role here. In this era of cutting-edge technology, deep learning (DL) is highly used in every sector. There are several existing methods to diagnose pneumonia but they have accuracy problems. In this study, an automatic pneumonia detection system has been proposed by applying the extreme learning machine (ELM) on the Kaggle CXR images (Pneumonia). Three models have been studied: classification using extreme learning machine (ELM), ELM with a hybrid convolutional neural network - principle component analysis (CNN-PCA) based feature extraction (ECP), and ECP with the CXR images which are contrast-enhanced by contrast limited adaptive histogram equalization (CLAHE). Among these three proposed methods, the final model provides an optimistic result. It achieves the recall score of 98% and accuracy score of 98.32% for multiclass pneumonia classification. On the other hand, a binary classification achieves 100% recall and 99.83% accuracy. The proposed method also outperforms the existing methods. The outcome has been compared using several benchmarks that include accuracy, precision, recall, etc.

by pneumonia due to their weak immune systems. Every year, many people die due to a lack of proper treatment because the patient either does not take it seriously or is identified late. It is the prime reason behind the death of children aged less than five years on a global scale and the death rate is around 95% in developing countries [1]. In 2015 around 17,850 (15%) died from pneumonia among the total deaths of 119,000 children who are aged five years or less in Bangladesh [2]. Around three children die every hour, 67 per day and 24,300 per year in Bangladesh [3]. Pneumonia is diagnosed using Chest X-Ray (CXR) images. It shows the spot of the infection and the extent to which the virus infected the lung. A radiologist investigates CXR images to identify the presence of pneumonia or any other kinds of lung diseases. However, for a human to diagnose and correctly recognize pneumonia using the CXR images is a time-consuming and quite challenging task. The CXR images are sometimes unclear and fuzzy, which drives an erroneous detection. To reduce the number of death, a trustworthy diagnosis system is required. In this modern technological era, machine learning (ML) performs a great role in automatically detecting diseases such as heart diseases, breast cancer, and brain cancer. CXR images can be used to train an ML model for developing a trustworthy automatic pneumonia detection system that will lessen the workload of radiologists. The novel contribution in this work: • Extreme learning machine (ELM) has been used to reduce training time cost as it has no iterative tuning parameter instead of traditional deep learning algorithms. • Image contrast has been enhanced by applying contrast limited adaptive histogram equalization (CLAHE) to achieve better accuracy. • Features have been extracted using new hybrid convolutional neural network -principle component analysis (CNN-PCA) to get discriminant features and reduce the model complexity. CNN has been used to extract features from raw data, which PCA follows. • Multiclass (viral pneumonia, bacterial pneumonia, and normal) and binary (pneumonia and normal) class classification has been performed using different machine learning approaches such as CNN, CNN-ELM, CNN-PCA-ELM, and CLAHE-CNN-PCA-ELM. The next section II describes recent research in this field. The proposed methodology of the model is described in section III. Section IV presents the results obtained from the ECP-CXR model, which is compared to the findings of other recent research. Finally, in section V, the key conclusions are presented.
Model 2, consisting of 3 convolutions and two dense layers, achieved the highest validation accuracy and recall of 92.31% and 98%, respectively. Chouhan et al. [5] proposed an ensemble model based on transfer learning for detecting pneumonia that achieved test accuracy and recall of 96.4% and 99.62%, respectively. The ensemble model consists of AlexNet, DenseNet121, InceptionV3, resNet18, and GoogLeNet neural networks. Rahman et al. [6] performed three experiments, with one of them is pneumonia classification. They trained four pre-trained models, including AlexNet, ResNet18, DenseNet201, and SqueezeNet, to carry out the experiments to detect pneumonia using CXR images. Among them, DenseNet201 achieved the highest accuracy of 98% for pneumonia classification. To train the models, images were resized and normalized according to the requirements of the transfer learning models. Image augmentation techniques (Rotation, Scaling, and Translation) were also performed. Wu et al. [7] proposed a hybrid model named ACNN-RF for pneumonia detection that achieved 97% accuracy. The full study can be divided into three steps. In the first step, an adaptive median filtering technique has been used to remove noise from CXR images. In the second step, a CNN model was developed to extract features. And, in the final steps, trained a Random Forest (RF) model with the extracted features for the detection process. Mittal et al. [8] used Integration of convolutions with capsules (ICC), Ensemble of convolutions with capsules (ECC), and EnCC (where n = 3, 4, 8, 16) on Mendeley CXR image dataset for pneumonia detection. Among them, E4CC achieved the optimal accuracy of 96.36%. Sarker et al. utilized bilateral filtering for noise reduction and enhanced the contrast of the CXR images using CLAHE [9]. Using a separable convolution technique and the benefits of deep residual learning, their suggested model reached a classification accuracy of 98.82%. Abin et al. proposed a dehazing approach for the detection of pneumonia from CXR images [10]. They removed the noise and improved the contrast of the CXR images using CLAHE. Ayan et al. [11] used transfer learning-based model Xception and VGG16 for pneumonia classification. VGG16 achieved 87% accuracy. Sharma et al. [12] used two CNN models, CNN with dropout layer and CNN without dropout layer, for detection purposes. They also used different augmentation techniques to avoid overfitting. The model with augmentation and dropout together achieved 90.68% accuracy. Heidari et al. [13] used transfer learning-based model VGG16 to detect COVID19 based pneumonia using X-ray images and achieved 98.8% accuracy. Liang et al. [14] used a combination of residual thought and dilated convolution to establish a system that can identify pneumonia. The proposed method achieved an f1 score of 92.7%. Ibrahim et al. [15] proposed AlexNet based pneumonia classification. The pneumonia dataset contains four classes: COVID 19, viral pneumonia, bacterial pneumonia, and normal pneumonia. They trained the model for different combinations such as 2-class, 3-class, and 4-class classification. Ibrahim et al. [16] performed multiclass classification that includes COVID19, pneumonia, and lung cancer. They used both the CXR and CT scan images to train four models for classification, including VGG19-CNN, ResNet152V2, ResNet152V2 + Gated Recurrent Unit (GRU), and ResNet152V2 + Bidirectional GRU (Bi-GRU). Lahoura et al. [17] proposed a breast cancer diagnosis system which was a cloud computing-based framework using ELM, and performed a comparison with different state-of-art models. ELM was shown a favorable result. Khan et al. [18] proposed a deep extreme learning machine (DELM) method for COVID19 detection, which achieved 97.59% accuracy. Many researchers have investigated pneumonia detection and have presented many machine learning approaches to address this issue. Some of these articles have presented lower true positives rates, which are not acceptable in the case of detecting pneumonia-affected patients. Again, most articles only consider binary class; it is also important to distinguish between bacterial and viral pneumonia. The majority of approaches failed to deliver the high accuracy necessary, particularly in 3-class scenarios. The existing methods consist of heavyweight models such as VGG, AlexNet, DenseNet121, Inception, ResNet, etc., which have a large number of tunable parameters, and hence they require a longer training time and higher computational power. In contrast, the proposed model consists of a simple CNN as a feature extractor. ELM has been used to classify binary and multiclass pneumonia as a classifier, and this does not have any iterative tuning parameter. This approach saves time for the training operation and reduces the classifier complexity.

III. METHODOLOGY
Data scientists have used several ML and DL architecture to predict different life-threatening diseases, such as COVID19, pneumonia, heart diseases, etc., using medical images in the last few decades. This study has attempted to correctly identify pneumonia from chest x-ray images by combining the benefits from both DL and ML algorithms. First of all, the chest x-ray images of pneumonia patients have been collected and subsequently enhanced the quality of the images using CLAHE. Since the images are different in size, all the images are resized into 224 × 224. The last part of the image pre-processing has been done by normalization. Then, a novel CNN model with six convolution layers has been designed for extracting the most discriminant features from the pre-processed images. Furthermore, the features have been standardized before PCA. Then PCA has been used for further selecting the essential features by maximizing the variance and also reduced the features that are not absolutely necessary [19]. Finally, an (ELM) model has been designed for the correct classification of the pneumonia patients. Fig. 1 shows a flowchart of the proposed methodology.

A. DATA DESCRIPTION
The CXR images of pneumonia were collected from Guangzhou Women and Children's Medical Centre, Guangzhou [20] which is also publicly available in Kaggle [21]. The database consists of 5857 chest x-ray images with the resolution varying from 400 pixels to 2000 pixels. The images from the database are separated into three categories: normal, bacterial pneumonia, and viral pneumonia. The samples of each class are shown in Fig. 2. The number of images from normal, bacterial pneumonia, and viral pneumonia patients are 1583, 2779, and 1495 respectively, which are shown in Table 1. Here two types of classifications have been performed where one is for multiclass classification, and the other is for binary class classification. For binary classification, bacterial and viral pneumonia have been combined into pneumonia class. Then the number of images for normal and pneumonia patients is 1583 and 4274, respectively. This implies that VOLUME -, -- normal class had been used to represent 27% of the images, while pneumonia had been seen in 73% of the images. The datasets have been divided into training and testing sets. For multiclass classification, the whole dataset has been divided into a ratio of 89.81 : 10.19 and for binary class classification, the whole dataset has been divided into a ratio of 90 : 10 for training and testing the model. The data splitting is shown in Table 2.

B. PRE-PROCESSING
Preprocessing of image data is critical because the quality of image preprocessing affects the classification results.

1) Image Contrast Enhancement
In this study, the image contrast has been enhanced by Histogram Equalization (HE) to improve the model accuracy. Contrast is determined by diversity in color or brightness with one object to other objects in a similar viewpoint. It makes one object recognizable from other objects. It is mainly used for images which are related to scientific works such as X-Ray, Satellite, and Thermal images [22]. As CXR images have been used here, it is also useful to apply image contrast enhancement for this study. An HE technique named Contrast Limited Adaptive Histogram Equalization (CLAHE) is used for this study. It is a version of Adaptive Histogram Equalization (AHE). HE method is applied for the entire image. In contrast, for AHE, an entire image is divided into small portions of rectangular-sized named tiles of an image, and AHE is applied for all the tiles individually [23]. It enhances the local contrast and edges. However, the AHE overamplifies the noise in the region where the image is nearly uniform [24]. CLAHE solves this problem by limiting the amplification. It clips the amplification by the clipping factor. Though the normal chest x-ray images are grayscale, the chest x-ray images used in this study are obtained from the Kaggle pneumonia image dataset, defined by 3 channels. CLAHE can be used for gray or color images [25]. In this study, the contrast has been enhanced by CLAHE while considering 'L', 'A', 'B' information [26], and the input data with three channels are required. For that reason, the input data is described as 3 channels instead of 1. For this study, a clipping factor of 2 and the size of a tile is 8 × 8 have been employed.

2) Resizing
The database contains images in various sizes, and they are converted to a specific size so that it is easy to fit these into the CNN model. As a result, after performing CLAHE on the images the images are adjusted to the image for the model to 224 × 224 pixels in size.

3) Normalization
A large number of intensity values are used to depict an image. Hence, normalization is carried out to avoid any complexity without a large number of image pixels. The scale is adjusted from 0 − 255 to 0 − 1 to lessen the complexity of the images by dividing image pixel values by 255.

C. HYBRID CNN-PCA BASED FEATURE EXTRACTION
Availability of a large number of data, i.e., high dimensional data, creates problems in developing an effective predictive model. Initially, the CNN model has been used to reduce the total (224 x 224) features to 512. Then PCA has been implemented to reduce the 512 features into 100.

1) Feature Extraction using CNN from Raw Data
The most crucial aspect of the classification problem is feature extraction because the performance of a model is determined by how well the important features from the chest x-ray images are extracted. It is necessary to extract the good features that have discriminated between the two classes to improve the classification performance of the model. Feature extraction is a technique that converts higher dimensional data into lower-dimensional, non-redundant, informative data [27]. It enables further data processing and improved data management. Hence, a novel deep CNN has been built for extracting 512 important features for pneumonia detection using chest x-ray images since the features for these images are more complicated. Fig. 3  demonstrates the proposed CNN model. A total of six convolutional layers form the proposed CNN model, whereas, after two consecutive convolutional layers, batch normalization and a max-pooling layer are used. Batch normalization has been used because it makes the model faster to run and more stable by re-centering and re-scaling the inputs of the layers [28]. A pooling layer between two successive convolutional layers has been placed. Max-pooling with 2 × 2 filters has been utilized, which select the largest value from each cluster's whole neuron at the convolutional layers and can extract the most crucial components from images [4], [29], [30]. A pooling layer between two successive convolutional layers has been placed. Max-pooling with 2 × 2 filters has been utilized to select the largest value from each cluster's whole neuron at the convolutional layers and extract the most crucial components from images. The 'SAME' padding has been introduced in the first two convolutions layers since the output has been calculated by applying the filters to all image tuples. As a result, the border elements have been checked because they can often include crucial features. The border elements were computed with zero paddings. The 'VALID' padding, on the other hand, ignored the border elements. ReLU has been employed as an activation function to avoid the problem of disappearing the gradient [4]. Two dropout layers with a probability of 0.5 have been utilized where one dropout is used after the final max-pooling layer, and another is used after the first fully connected layer. Here dropout is used for reducing overfitting by frequently avoiding training all nodes in each layer during the training phase, resulting in a significant boost in training speed [31]. The Adam optimizer has been chosen because it is highly accurate for CNNs, performs better when training on large amounts of data [32]. As the CXR images dataset used in this study is labeled, the loss has been calculated using sparse categorical cross-entropy with the provided label in the dataset. The autoencoder, generally used for unlabeled data, has not been considered in this study [33]. The learning rate of the proposed model is 0.001, and trained the model for 100 epochs with a batch size of 32. Finally, the last dense layer has been used to extract the 512 discriminant features from each image. Table  3 shows the summary of the deep CNN model.

2) Feature Extraction using PCA
PCA is an unsupervised statistical technique of dimensionality reduction. It transforms the possibly correlated features into linearly uncorrelated features [34]- [36]. It can be described using the following steps: 1) Perform data standardization which means the values of each feature in the data are given a zero-mean and unit-variance through feature standardization [37]. It has a significant effect on PCA. as scaling can affect the covariance matrix. To determine the correlation between two features, element-wise multiplication between them is performed. Hence, if both features are not scaled with the same range, the purpose of the covariance matrix will be diverse. After extracting 512 VOLUME -, -- features using CNN, standardization, i.e., feature scaling, has been performed on the data. It has been performed by this using the equation 1 [38].
Where, x = Original feature vector x = Mean of that feature vector σ = The standard deviation Let, Z be the standardized version of feature matrix X.
2) The covariance matrix for this study has been determined using the matrix A = Z · Z T . 3) After calculating the covariance matrix, the eigenvector and the eigenvalue have been calculated. Calculate eigenvalues using (A − λI) = 0 and eigenvector by solving (A − λI)V = 0 for different values of lambda. For this study, there are 512 eigenvalues, and each eigenvector consists of 512 elements. 4) Then sort the eigenvalues in descending order and the respective eigenvector, and among the 512 eigenvalues, top 100 eigenvalues are selected, and the dimension of the eigenvector matrix, E is (512 × 100). 5) Multiply the feature matrix, X with the eigenvector matrix, E. The dimension of the new feature space is (5856 × 512) · (512 × 100) = (5856 × 100). Hence, the dimension is reduced.

D. EXTREME LEARNING MACHINE
ELM has been proposed by Huang et al. [39] to minimize the training time costing due to the iterative model parameter tuning process. In this study, for binary class classification, 100 features have been fed into the input nodes, and there are 500 nodes in hidden layer and one output node. For threeclass classification, there are a total of 3 output nodes in the ELM, which is shown in Fig. 4. As simple single hidden layer ELM has been used, there is no error back-propagation process. Output weight matrix has been calculated using the pseudo-inverse technique. Let's, consider a training sample (1,m) , y (1,t) : m R+, t R+}. X be the input and Y be the output for this study. Then, all the train samples can be represented by matrix format such as (1,m) . . . . . . . . .
x (n,1) x (n,2) · · · x (n,m) Where, n be the number of instances, m be the number of attributes and t be the number of output nodes. Hence, for this study m = 100 as the feature space has been reduced to 100 and value of n will be changed with the size of train and test dataset. For binary class t = 1 and for multiclass classification t = 3. Then, randomly generates the input weight W (m,N ) and bias B (1,N ) matrix. Where, N =number of hidden nodes. N ) . . . . . . . . .
Where, m be the number of attributes and N be the number of hidden nodes. For this study, the value of N will be 500. Then, calculate the output of the hidden layer using H (n,N ) = G(X (n,m) · W (m,N ) + B (1,N ) ). Where, G(x) be the activation function. In this study ReLu activation function has been used. N ) . . . . . . . . .
Finally, calculate the output layer weight β (N,t) by using the Moore-Penrose pseudo inverse.

A. QUANTITATIVE ANALYSIS
A confusion matrix is a tabular representation in an easily understandable format used to analyze the performance of a classifier. Table 4 represents the confusion matrix. Where true positive, true negative, false positive, and false negative represented as TP, TN, FP, and FN respectively. For evaluating the classification performance of the different approaches several evaluation criteria have been considered which are accuracy, precision, recall, sensitivity, specificity, f1-score, receiver operating characteristics (ROC) curve [40]- [43].

B. RESULT ANALYSIS
The experiments and outcomes of various performance measurements are mentioned in this section. Keras had been used to implement all of the machine and deep learning algorithms, with TensorFlow as the backend, which runs at the Pycharm Community Edition19 (2020.2.3 × 64) software. For the training and testing of the model, a PC with an Intel(R) Core(TM) i7-6700 CPU @3.40GHz processor, 32GB RAM, and an NVIDIA GeForce GTX 1650 SUPER 4 GB GPU running on a 64-bit Windows 10 Pro operating system has been used. For the training and testing of the model, a PC with an Intel(R) Core(TM) i7-6700 CPU @3.40GHz processor, 32GB RAM, and an NVIDIA GeForce GTX 1650 SUPER 4 GB GPU running on a 64-bit Windows 10 Pro operating system has been used. First, the chest x-ray images have been preprocessed. After that, a novel CNN model for dimensionality reduction has been proposed. Fig. 5 showed the training and validation accuracy for multiclass is 98.28% and 82.98% respectively, whereas the training and validation loss is 0.015 and 0.52 respectively during the dimensionality reduction process. Fig. 6 demonstrated the training and validation accuracy, loss during the features extraction process for binary class. The training and validation accuracy is 99.98% and 99.78%, respectively, whereas the training and validation loss is 0.0027, and 0.10 respectively.

1) Results for Multiclass Classification
The extreme learning algorithm has been trained using 5260 data of both normal, bacterial, and viral pneumonia-infected VOLUME -, -- persons where 1421 images were from normal, 2496 images from bacterial pneumonia, and 1425 images from viral pneumonia patients. To calculate the performance of the model 597 data, we have tested our ELM model where the number of images for normal, bacterial, and viral pneumonia patients is 162, 283, and 152. First, all 512 features have been used to train the model. Second, PCA has been implemented for selecting 100 discriminant features to check the performance of the model increasing or not. Finally, CLAHE has been used to preprocess the images, and then PCA has been used that achieved the best result in every benchmark. To evaluate the stability of the ELM model, a confusion matrix has been employed to calculate the accuracy, precision, recall, f1-score, and area under the curve (AUC), which are shown in Fig. 7 A, B, and C. In the medical sector, the recall should be maximized since the patient who has been infected with different types of pneumonia must be correctly identified as pneumonia. For detection pneumonia CNN provides an average precision, recall and accuracy of 0.82, 0.80, and 82.25%, respectively which are shown in Table 5. It was observed that the classification performance using CNN only was not quite good specially for multiclass classification. For that reason, different combinations have been explored for the classification of multivariate pneumonia.  Table 6 showed the result for classification of pneumonia using ELM-CNN without CLAHE & PCA. That means that after extracting features using CNN, these three classes have been classified with the help of ELM. The model has achieved an accuracy and recall of 97.09% and 96%. Then, the model's performance has been enhanced by applying PCA to select 100 features from the 512 features, which are extracted using CNN. Then ELM has been used to classify pneumonia from these 100 features and achieved accuracy and recall of 97.98% and 97%. Table 7 showed the performance of the ELM using PCA for multiclass. Further, the performance has been improved by using CLAHE, which means enhanced the contrast of the images, extracted the feature using hybrid CNN-PCA, and classified the disease using ELM. Table 8 showed that the result with    accuracy and recall of 98.32% and 98%. The AUC of the ELM model for multiclass classification using only ELM, ELM with PCA, and ELM with both CLAHE & Hybrid CNN-PCA showed in Fig. 8 A, B, & C respectively. It is used as an evaluation measure to fine-tune the performance of a machine learning model [44].
From Table 9, it is concluded that the training time utilizing the ELM with hybrid CNN-PCA is just 15.977 seconds longer than the training time using the CNN only. Following the extraction of the features, the testing time for the ELM using hybrid CNN-PCA is just 0.00099 seconds, which is significantly less than the testing time of 1.377 seconds using the conventional CNN. It is negotiable if the classification performance of the ELM with hybrid CNN-PCA has been compared to the classification performance of the CNN-based system. As a result, it can be argued that the overall performance of ELM has been significantly superior to the performance of the CNN alone. From Fig. 9, it was observed that the accuracy and specificity of ELM with both CLAHE and CNN-PCA was  found to be 98.32% and 98.9%, respectively, which were higher than that by ELM without PCA (97.09% and 98.6%), ELM with only PCA (97.98% and 98.8%) and CNN only (82.25% and 90.10%). The image contrast has been improved by applying CLAHE that makes the images more classifiable. The hybrid PCA-CNN feature extraction technique reduces the number of features, removing irrelevant and redundant features and minimizing the model parameter. Due to these reasons, the ELM model with CLAHE & PCA has produced better results than the other three models for multiclass classification.

2) Result for Binary Classification
For binary class classification, 5271 data has been used for both normal, pneumonia-infected images to train the model, where 1425 images were from the normal patients and 3846 images from the pneumonia patients. To calculate the model's performance, 586 data was used to test the ELM model, where 158 images were from normal and 428 images from the pneumonia patients. Here the same approaches have been used as multivariant classification. Fig. 10 A, B, C show the confusion matrix. The precision, recall, f1-score have been calculated from the confusion matrix. Table 10 shows the result for binary classification using only CNN.
The average precision, recall, and accuracy is 0.94, 0.92, and 94.53%, respectively. The average precision, recall, and f1-  Table 11. The model has been achieved precision, recall, and f1score of 0.99, 0.99, and 0.99, respectively, using ELM with Hybrid CNN-PCA. Table 12 shows the result. In the final experiment, CLAHE and Hybrid CNN-PCA both have been used. The proposed model has achieved precision, recall, and f1-score of 1.00, 1.00, and 1.00. It has shown a promising result compared to the existing methods. The ROC curve of the ELM model is shown in Fig. 11. From Table 14, it is observed that the training time using the ELM with hybrid CNN-PCA is only 19.632 seconds longer than that by the CNN only. After extracting the features, the testing time for the ELM with hybrid CNN-PCA is only 0.00498, which is much smaller than the testing time of 1.340 using the CNN only. It is negotiable if the classification performance of the ELM with hybrid CNN-PCA has been considered against the performance using CNN. Therefore, it is concluded that the overall performance of the ELM for the binary classification of pneumonia is relatively better than the performance of the CNN alone. From Fig. 12, it was observed that the accuracy and specificity of ELM with both CLAHE and CNN-PCA was found to be 99.83% and 100%, which were higher than the accuracy of ELM without PCA (95.90% and 97.66%), ELM with only PCA (99.48% and 99.53%) and CNN only (94.53% and 96.96%). Due to similar reasons for multiclass classification, the ELM with CLAHE CNN-PCA performed better than the other three approaches for the binary class classification. In multiclass classification, the pneumonia patients were additionally divided into viral and bacterial pneumonia. As the number of classes increased, the model performance decreased.

C. PERFORMANCE COMPARISON ANALYSIS
From the results, it was clear that the accuracy of the binary classification was higher than that of the multiclass classification. It should be acknowledged that it is more challenging to increase the accuracy as well as other performance criteria for multiclass classification, particularly after reaching a certain degree of accuracy. In the case of multiclass classification, the accuracy was raised by 0.89% when compared between the ELM without  CLAHE & PCA (Table 6) and ELM with hybrid CNN-PCA (Table 7). Again, a further increase in accuracy of 1.23% was observed when compared between the ELM without CLAHE & PCA (Table 6) and ELM with both CLAHE & hybrid CNN-PCA (Table 8). It is expected that this incremental accuracy would help in the detection of these types of diseases. Furthermore, the accuracy of the three-class classification using only CNN was 82.25% (Table 5), while after using ELM, this classification accuracy was boosted up to 98.32% (Table 8), which is almost 16% higher than that obtained by the CNN only. For three-class classification using ELM, better classification performance was also obtained than the state-of-the-art models, which are shown in the latter section. In this section, the proposed approaches have been also checked for statistical significance. In the CNN-ELM approach, first, CNN is used for features extraction and ELM for classification. On the other hand, in the CNN-PCA-ELM approach, the hybrid CNN-PCA is used for features extraction and ELM for classification. Here, these two main approaches have been checked for both binary and multiclass classifications. To choose the best model among the candidate models, scores on different evaluation criteria such as accuracy, sensitivity, specificity, etc., are used. The best model has been chosen from the direct comparison through a statistical significance test (t-statistics) [45]. P-values have been calculated for both the CNN-PCA-ELM and CNN-ELM approaches and binary and multiclass classifications to identify any significant difference [46], [47]. P value of 0.031 (P<0.05) was obtained when compared between the CNN-PCA-ELM, and CNN-ELM approaches, indicating that the CNN-PCA-ELM approach could produce the best prediction accuracy.

D. PERFORMANCE COMPARISON WITH PREVIOUS WORKS
In this section, it has been demonstrated that how the extreme learning machine method performs better with PCA than earlier research in this sector. In section II, the proposed method has already gone over the specifics of the previous works. Rahman used four different pre-trained deep CNN transfer learning models [6]. They considered both multiclass and binary class classification. For multiclass classification, they used 597 data for testing their model and achieved an accuracy of 93.3%. The same data has been used for testing the ELM with CLAHE & PCA and achieved a higher result with an accuracy of 98.32% which are shown in  [14]. Finally, they achieved an accuracy of 90.5%, a recall of 96.7%, and an AUC of 95.3%. Jain et al. used the same dataset for training and testing their various models using 624 images [4]. Their models achieved the best recall and accuracy, at 98% and 92.32%, respectively. Using VGG16 over the Xception network, Ayan et al. had the best accuracy [11]. They tested their model using 624 chest x-ray images and achieved accuracy and recall of 90.5%, and 96.7% respectively. Chouhan et al. utilized the same dataset for training their ensemble model, and they got 96.39% of accuracy, 99.62% of recall, and 99.34% of AUC [5]. The same data has been used for training He has over 150 refereed publications including journal (18 Q1 ranked in Scimago) and conference papers with h-index 20, i10-index 33, and 1205 citations. These cover the target sectors: "Energy and Mining Technology", "Quantum Information, Advanced Digital, Data Science and ICT", and "Cyber Security". and it is in these areas that he is renowned for his work both Nationally and Internationally justified by his research outputs. He has been supervising research associates and PhD Students in the above areas. He has published and presented over 100 technical papers in international journals, conferences and books. He is acting as the Executive Editor for an International Journal published by Taylor and Francis (Advances in Materials and Processing Technologies). He has received funding from Innovate UK to conduct Eight Knowledge Transfer Partnership projects with industry worth more than a million.