Dual deterministic model based on deep neural network for the classification of pneumonia

Abstract

Pneumonia is a disease caused by the virus (flu, respiratory Syncytial Virus) or bacteria. It can be fatal if not diagnosed and treated at an early stage. Chest X-rays have been widely utilized to diagnose such abnormalities with high exactitude and are primarily responsible for the augment real-world diagnosis process. Poor availability of authentic data and yardstick-based approaches and studies complicates the comparison process and identifying the safest recognition method. In this paper, a Dual Deterministic Model (DD-M) is proposed based on a Deep Neural network that would identify Pneumonia from chest X-ray and distinguish the cause in case of either viral or bacterial infection at an efficiency equivalent of an active radiologist. To accomplish the automated task of the proposed algorithm, an automatic computer-aided system is necessary. The proposed algorithm incorporates deep learning techniques to understand radiographic imaging better. The results were evaluated after implementing the proposed algorithm where; it reveals various aspects of the chest infected with Pneumonia compared to the healthy individual with approximately 97.45% accuracy and distinguishes between the viral and bacterial infection with the efficiency of 88.41%. The proposed algorithm with an improved image dataset will help the doctors diagnose.

Keywords

Viral Pneumonia Bacterial Pneumonia deep learning dual deterministic model (DD-M)convolutional neural network (CNN)

1. Introduction

Pneumonia is one of the soul frightening, extensive infectious diseases that can be triggered through the entering of bacteria, viruses, or fungi in the body of a human being that requires a precise diagnosis at an early stage for a suitable treatment [1]. In the case of severity, it can cause death. To diagnose Pneumonia, doctors and health care specialists use a chest X-ray that is proved to be an excellent imaging sense modality. However, it is possible to spot Pneumonia using other non-contact procedures such as Computed Tomogram (CT) images and High-Resolution Computed Tomography (HRCT) [2]. Meanwhile, precise assessment of the X-ray images is a significant aspect for the radiologist having proficiency and capability [3]. As it is an approximation by the World Health Organization (WHO), approximately two-thirds population of the world is impotent to approach the radiologist for precise diagnosis [4]. An automated computerized system is required to calculate the X-ray images that help the doctor and healthcare facilitator diagnose can be proved very useful. In such cases, researchers have adopted image processing algorithms to detect Lung abnormalities [5]. For the classification of Pneumonia, a chest X-ray is employed as an imagining sense modality for the discovery of the disease through a trained radiologist. A recent study illustrates the diagnosis of Pneumonia using a deep learning framework for automated recognition of Pneumonia and other diseases on X-ray images to support the health consultants [6]. Furthermore, an immediate diagnosis of Pneumonia type along with proper medication may be helpful in saving the patient’s life. Machine learning-based multiple-piece explanations have also been endorsed for extracting valuable characteristics to mechanize the process of categorization. Over time medical sciences progressively expand, and Chest X-ray images are the best method to diagnose Pneumonia in the patient [7].

In most cases, X-ray images developed by various means are not very clear and provide noticeably less information due to which they are sometimes misdiagnosed, and the same can be said for the Pneumonia categories such as viral or Bacterial Pneumonia. Due to misdiagnosis, wrong medication may be offered to the patient, deteriorating their condition [8]. Development in methods allows researchers to evaluate the results more precisely. Health check personnel get a dissected picture by adopting a vast algorithm that certainly influences the correctness of diagnosis outcomes. In third world countries, many irregularities are reported to diagnose Pneumonia due to the deficiency of skilled radiologists [9]. A structure that can classify Pneumonia and its type is essential, which may help the concerned radiologists promptly diagnose the type of Pneumonia after getting hold of an X-ray image.

Deep learning models, such as CNN, fascinated many researchers to examine medical records because they are altered and made for extensive data analysis. Deep learning models can pick up the features pulled out from the image database utilizing multiple layers of filters [10]. Since the beginning of the deep learning algorithm in the medical field, it is becoming more robust and widely held among researchers due to its automatic diagnostic feature and precision [11]. It has been brought into being from the previous studies done by researchers Cernazanu and Holban. The segmentation of the Chest X-ray has been defined through a convolutional neural network [12], the segmentation of images into the bone and non-bone tissue was elaborated briefly. The primary purpose of their study was to build a spontaneous methodology that would be able to detach the bone tissue from the whole image of the chest X-ray through a convolutional neural network, which the objective of their study would achieve. The CNN works on the principle of analyzing the pixels having a graphic picture that have been later categorized into the bone and non-bone tissue [10, 12]. Biomedical applications such as cancer finding in breast, artery-clogging, and heart diseases consume Artificial Intelligence for their proper diagnoses [13, 14]. Srivastav, Devansh et al. presented a technique based on Generative Adversarial Network (GAN) to generate synthetic image dataset via augmentation and applied transfer learning algorithm to improve the accuracy with modified dataset [15]. Researchers also included the concept of Region of Interest (ROI) to categorize Pneumonia and its type using a chest X-ray image dataset [16]. By applying a similar technique to further research, the classification of lung cancer has been done that introduced an image processing technique [17] along with its application to explore whether cancer is malicious or else at its initial stage. Deep learning algorithms require an enormous amount of data to learn, which are easily accessible for various applications to train different deep learning models [18, 19]. Many researchers involved in developing deep learning algorithms for the classification of Pneumonia type have experimented in varying the deep learning parameter; in some cases, pre-trained classifiers are also augmented and applied to classify Pneumonia and its types [20, 21]. Through such advancement, diameter, perimeter, irregularity in categorization was automatically recognized. The techniques created on machine learning using CNN for the progression and development of the diagnosis of images [22]. To some degree, the machine learning algorithms are hard-hitting because assailants at any time can produce some models that cannot be trained by machines [23]. Since COVID-19 became a global phenomenon, few studies have been introduced on Artificial Intelligence to spot and classify it. Pre-trained models such as ResNet50, ResNet101, ResNet152, and Inception ResNet-V2 are extensively utilized for COVID-19 detection [24]; the study was accomplished using four different classes, in which the amount of ‘COVID’ images are 341, 2800 ‘Normal’, 1493 ‘Viral’ and 2800 ‘Bacterial’ were used.

Vikash et al. has used the transfer learning technique by incorporating ImageNet models to classify Pneumonia and achieved an accuracy of 96.39% [42]. Similarly, Saraiva, Arata Andrade used Artificial Neural Network to classify Pneumonia and attain an accuracy of 94.5% [46]. Khatri et al. also presented a similar technique for identifying Pneumonia type (viral or bacterial) by incorporating the EMD model and achieved an accuracy of 83.3% [47]. Gu et al. used a custom VGG16 model, and their work was distributed in two parts; they used Fully Convolutional Network (FCN) to classify the Lung region affected and used Deep Convolutional Neural Network (DCNN) to classify the category of Pneumonia [48].

In this paper, a (DD-M) is projected to classify Pneumonia type in two levels. Level 1 used a (DCNN) model to classify an inputted X-ray image as either Normal or Pneumonia and the proposed model achieved an accuracy of 97.45%. Level 2 classification is only performed when Level 1 classifies an X-ray image to be Pneumonia. It identified Pneumonia type (viral or bacterial) using a custom DenseNet201 model and achieved an accuracy of 88.41%. Model is distributed in two different levels to achieve better accuracy as compared to the methods used by Khatri et al. [47] and Gu et al. [48], where the final layer of the model has three output/classes, for example, Normal, Viral, and Bacterial, while the planned model has two courses for each of its level and in both cases, it performed better.

2. Deep learning models

Deep learning is an interaction through which a machine can learn and retain a particular sort of information the way a human being does. In light of Deep learning, such algorithms are utilized for predictive analysis, which can automate them.

2.1 Convolutional Neural Network (CNN)

Figure 1.

CNN architecture.

A Convolutional Neural Network (CNN) is a well-known deep learning algorithm that recognizes and classifies elements in illustrations for computer visualization [25]. Working with images, a convolutional neural network is the most acceptable option. This system acquired inspiration from the visual cortex, just similar to the brain of the human processing the image of something after visualizing it [26]. The brain processes an enormous volume of information as soon as the visual cortex views anything. Each neuron adds to its capacity belonging to its related field, and they communicate with other neurons to deliver their specific information and cover the whole area of view [27, 28]. Neurons in CNN are like the neurons existing in the brain, and both operate with their respective field of view. Information processing covers the entire image through each node, as given in Fig. 1, which expresses CNN’s architecture. Each layer is organized to be capable of identifying the same patterns such as edges, curves, peaks, etc., using their detailed kernel matrix [17, 29]. Three layers in CNN are usually used in its architecture: convolutional layer, pooling layer, and a fully connected layer. Several convolutional layers or multiple pooling layers in the CNN model, for each layer output, is forwarded to its following adjacent layer. Since it is a multilayered model, the convolutional and pooling layers are mainly hidden layers, and Fully Connected Layer (FCL) is for the classification [16].

Figure 2.

Feature extraction using kernel matrix in convolutional layer.

The convolutional layer fills in as the fundamental building block of the CNN architecture, which bears the vast majority of the computational load inside the CNN network. CNNs are feedforward Neural Networks limited by two necessities, neurons inside the selected field latches with similar features to diminish the complexity and hold the spatial structure and its weights. The principal function of the Convolutional layer is to calculate the Dot Product (DP) of the kernel matrix [21] and a specific percentage on the image field as disclosed in Fig. 2, and this percentage is denoted as a window of $n\times n$ dimension. The size of the kernel is lesser than the size of an image, and it slides across the whole field following the track laid by the rows and columns, which form an opening of the same size as a kernel. The whole convolutional process is a two-dimensional illustration of image space, also denoted as an activation map that makes the kernel’s calculated response at every spatial location within an image [18]. Equation (1) indicates finding DP in the convolution block. The second hidden layer is a pooling layer that summarizes the dot product values calculated using pooling functions such as L2 norm of the neighborhood, max pooling, etc. The fully connected layer mainly bridges the gap between the input and the output. It utilizes non-linear functions such as sigmoid and Rectified Linear Unit (ReLU) to cater for the non-linearity presented by images. Usually, ReLU is used by researchers for its junction.

It is established on the multi-layer neural network developed to evaluate graphical inputs and accomplish chores such as image categorization, dissection, and target revealing, benefiting autonomous vehicles [30]. It can be summed up that the development, improvement, and augmentation of images utilizing image histogram equalization, deep neural network, and convolutional neural network upholds in achieving a sophisticated response rate and error; nevertheless, it requires farseeing training stretch that would be due to the development that required too much time [31].

$\displaystyle\textit{Conv2D}_{(x,y)}^{t}=\sum_{y=1}^{N}\left(f_{(x,y)}*k_{(x,y% )}^{t}\right)$ (1)

where,

•

$k$ : Kernel Matrix

•

$(x,y)$ : Image rows and columns, respectively

•

$f$ : Current Frame

•

$N$ : Number of iterations for DP

•

$t$ : Kernel Number

2.2 Pre-trained classifiers models for transfer learning

To authenticate the suggested DDM algorithm, three State of the Art (SOTA) networks are used. AlexNet [32] and ResNet [37] are the two pretrained algorithms used. A short overview of the two pre-trained networks talked over in Sections 2.2.1 and 2.2.2.

2.2.1 AlexNet

Figure 3.

AlexNet architecture.

AlexNet is amongst the most well-known pre-trained classifiers available for classification [33]. It has a class vector of 1000, implying it can classify 1000 different classes. AlexNet calls for an input of an image with three channels or a Red Green Blue (RGB) image with the resolution of 227 $\times$ 227. It utilizes 5 convolutional layers, 2 FCL, and a softmax layer [34, 35]. Additional details can be visualized in Fig. 3.

2.2.2 ResNet18

Residual Network or ResNet was introduced to cater to two subjects: degradation problem and vanishing gradient [36, 37, 38]. Suppose the convolutional layer is added with a more precise activation function. In that case, the gradient of the loss function falls to zero, in turn rendering the network particularly difficult to train [38] without error. ResNet18 is not the only alternative of its kind, ResNet50 and ResNet101 being the other two. Names of all three are created on the number of layers they hold [39]. Alternatives of ResNet are widely used in medical imaging [40, 41] and its related classifiers. ResNet18 is inducted to identify Pneumonia in this paper for fair analysis of the proposed model. ResNet focuses on the residual reply between the sample dataset to learn rather than learning their features levels. Figure 4 shows the architectures of ResNet18.

Figure 4.

ResNet18 architecture.

3. Dataset and augmented dataset used

The well-trained model always gives precise results depending on the program that feeds to the learning system of the model. In most cases, the picked-up data set is not up to the mark and fails to achieve the new information of the sample corresponding to the annotation. A few pre-processing steps included training the model and applied transformation, such as moving, zooming, and suppressing the noise. This way, new sample information could be improved in the applied model.

Figure 5.

Augmented dataset for Pneumonia (row 1) and Normal (row 2) categories.

The revolving action done for picture augmentation typically takes place by revolving the provided picture clockwise from a point ranging 0 - 360 degrees. It turns the picture pixel and fills the picture where pixels are gone from the picture. The scaling activity is used for amplification or reduction of the size of the picture, which is another technique of using augmentation. Altogether, 10% is done for the picture amplification. Picture modification is likely by either deciphering the picture in an even, vertical course or in both directions. Few samples of the augmented data set of the similar image are displayed in Fig. 5. In Table 1, a detail of the X-ray image dataset is given in which some images were casually augmented to enhance their quality, as displayed in Fig. 5. The dataset is distributed in four diverse classifications for two classifications: Normal, Pneumonia, Viral Pneumonia, and Bacterial Pneumonia.

Table 1

Total number of X-ray images used in the dataset

Normal vs Pneumonia (Level 1)	Quantity of X-ray images with augmentation (training set)	Quantity of X-ray images with augmentation (test set)
Normal	4000	400
Pneumonia	4000	400
Total	8000	800
Viral vs Bacterial (Level 2)	Quantity of X-ray images with augmentation (training set)	Quantity of X-ray images with augmentation (test set)
Viral Pneumonia	3000	350
Bacterial Pneumonia	3000	350
Total	6000	700

4. CNN based dual deterministic model (DDM)

To check the existence of Pneumonia through chest X-ray gives a workable explanation of conceivable outcomes through a deep learning algorithm. The model learns all the fundamental features corresponding by itself. Gradient descent has been used as a programming optimization that enhances the neurons or nodes until it finds the function of local minimum. For executing the projected DDM algorithm, an intel core i7 machine bearing 32 GB of Random-Access Memory (RAM) and Nvidia GeForce 1660 Super Graphics Processing Unit (GPU) is used. The algorithm is established on the python platform using the Keras library. The DDM model has two different levels, where it is trained for two separate identifications. The first level of classification is to identify the positive or negative case of Pneumonia, and if the case is identified as positive, only then it will hop to the second level for the classification between Viral and Bacterial Pneumonia. The process of the DDM model is demonstrated as a flow control process in Fig. 6. The classification is separate for their particular classification process to gain additional accuracy.

Figure 6.

Flow control DDM for testing.

After the system initialization, an X-ray image input is acquired and transferred for pre-processing. Since the proposed DDM requires an input array of 224 $\times$ 224 and a single channel image information (Grayscale image), which implies 224 $\times$ 224 $\times$ 1 resolution for an input image dataset, the acquired X-ray image is transformed to meet the model requirements for training or testing. Level 1 classification will identify the positive or negative case of Pneumonia, in case it is classified as normal, the proposed framework will exit the current Level 1 loop and wait for another input X-ray image, but if it predicts Pneumonia, the image information is delivered to Level 2 for classification from where it’ll distinguish the case of viral or Bacterial Pneumonia. A customized DenseNet201 model is utilized for the suggested model, which takes three channel inputs rather than one, so the previous image dataset was used without transforming each image into grayscale for Level 1 classification.

Table 2 represents the Level 1 classification in which it classifies the image as Pneumonia or normal. It has three convolutional layers activated using the ReLU activation function and three pooling layers. The total trainable parameters of the whole network are 498882. Since the variation within the dataset for the classification of Pneumonia or normal is quite large, a small model such as Level 1 classifier is enough to achieve approximately 97.5% Validation Accuracy.

Figure 7 represents the Level 2 classification in which it tends to classify the already predicted Pneumonia into Viral or Bacterial. For Level 2 classification, Customized DenseNet201 Model is employed with total trainable parameters of 558434 as displayed in Table 3 while attaining the approximate validation accuracy of 88.41%.

Table 2

Level 1 classification (Normal or Pneumonia)

Layers	Nos of filters	Filter size	Activation function	Padding	Stride	Feature map size	Trainable parameters
Input	–	–	–	–	–	224 $\times$ 224 $\times$ 1	–
Conv2D (1)	32	5 $\times$ 5	ReLU	–	2	110 $\times$ 110 $\times$ 32	832
Max-Pool (1)	–	3 $\times$ 3	–	1	1	108 $\times$ 108 $\times$ 32	–
Dropout	–	–	–	–	–	108 $\times$ 108 $\times$ 32	–
Conv2D (2)	128	3 $\times$ 3	ReLU	1	2	54 $\times$ 54 $\times$ 128	36992
Max-Pool (2)	–	3 $\times$ 3	–	1	1	54 $\times$ 54 $\times$ 128	–
Dropout	–	–	–	–	–	54 $\times$ 54 $\times$ 128	–
Conv2D (3)	256	3 $\times$ 3	ReLU	1	1	54 $\times$ 54 $\times$ 256	331968
Max-Pool (3)	–	3 $\times$ 3	–	1	1	18 $\times$ 18 $\times$ 256	–
Dropout	–	–	–	–	–	18 $\times$ 18 $\times$ 256	–
Flatten	–	–	–	–	–	82944	–
FCL (2)			Sigmoid	–	–	2 $\times$ 1	165890
Total	–	–	–	–	–	–	498882

FCL: Fully Connected Layer; Conv2D: 2D-Convolutional layer; ReLU: Rectified Linear Unit.

Table 3

Level 2 classification trainable and non-trainable parameters

Classifier	Total Parameters	Trainable Parameters	Non-trainable Parameters
DenseNet201 (regular)	18510146	188162	18321984
Customized DenseNet201	18880546	558434	18322112

Figure 7.

Level 2 classification using customized DenseNet201 (viral or Bacterial Pneumonia).

To authenticate the results of the trained model, four different performance parameters were utilized. Performance is assessed based on Accuracy, Sensitivity/Recall, Precision, and F1 score. Two pre-trained models (AlexNet and ResNet18) were also used to train the acquired dataset and assist as a reference to compare with the proposed model. For Level 2 classification, a comparison is performed between a regular DenseNet201 model with the customized DenseNet201 model. The performance is compared based on the specific dataset used to train the models; the performance of the suggested model may differ for other applications or any other dataset that may vary according to the application. The equations of the above-discussed Performance parameters are expressed in Eqs (2)–(5)

$\displaystyle\textit{Accuracy}=\frac{(\textit{TP}+\textit{TN})}{(\textit{TP}+% \textit{FN})+(\textit{FP}+\textit{TN})}$ (2) $\displaystyle\textit{Sensitivity / Recall}=\frac{(\textit{TP})}{(\textit{TP}+% \textit{FN})}$ (3) $\displaystyle\textit{Precision}=\frac{(\textit{TN})}{(\textit{FP}+\textit{TN})}$ (4) $\displaystyle\textit{F1 score}=\frac{(2*\textit{TP})}{(2*\textit{TP}+\textit{% FN}+\textit{FP})}$ (5)

5. Results

The algorithm’s performance used for testing and training with the help of a confusion matrix is given in Table 4. The performance of the DDM model using four performance parameters is given in Table 5, where the proposed DDM model for Level 1 (Normal or Pneumonia) classification performed better as compared to the pre-trained classifiers as it produces the highest accuracy as compared to the rest of the two pre-trained classifiers which was found to be 97.45%. In the case of Level 2 (Viral or Bacterial) classification, a pre-trained DenseNet201 classifier is used and customized for the specific task to obtain the best performance compared to the regular DenseNet201 model. Since DenseNet201 produced the best performance compared to the rest of the pre-trained classifiers for Level 2 classification, which was found to be 85.977%, it was selected for the customization. The customized version of the DenseNet201 model attains 88.41%, as displayed in Table 5.

Table 4
Confusion matrix of AlexNet, ResNet18 and DDM along with their respective accuracies

	AlexNet (Level 1)			AlexNet (Level 2)
True class	Predicted class			Predicted class
	N	203	11	V	180	12
	P	15	234	B	13	220
	ResNet18 (Level 1)			ResNet18 (Level 2)
	Predicted class			Predicted class
	N	271	14	V	241	15
	P	17	315	B	18	292
	(DDM) (Level 1)			DDM (customized DenseNet201 – Level 2)
	Predicted class			Predicted class
	N	197	2	V	171	7
	P	5	227	B	5	212

N: Normal; P: Pneumonia; V: Viral Pneumonia; B: Bacterial Pneumonia; DDM: Dual deterministic model.

Table 5

Performance parameters of DDM model (validation/testing accuracy) vs pre-trained classifiers

Algorithms	Accuracy (%)	Recall (%)	Precision (%)	F1 Score (%)
AlexNet (Level 1) (Normal or Pneumonia)	93.75	94.2	93.36	93.677
AlexNet (Level 2) (Viral or Bacterial)	83.06	83.359	82.72	82.95
ResNet18 (Level 1) (Normal or Pneumonia)	94.882	94.54	95.16	94.923
ResNet18 (Level 2) (Viral or Bacterial)	82.47	83.11	82.204	82.379
DDM (Level 1) (Normal or Pneumonia)	97.45	97.9	97.55	97.731
DenseNet201 (Level 2-regular) (Viral or Bacterial)	85.977	84.98	85.326	85.101
DDM (Level 2-customized DenseNet201) (Viral or Bacterial)	88.41	87.91	87.836	88.242

Figure 8 represents the graphical representation of the accuracy and loss for Level 1 classification. In the given scenario, a custom callback is utilized to achieve the best model that attains an accuracy of 97.45%. Figure 9 displays the testing result of the proposed DDM Level 1 classification model, identifying all six images without any error. Figures 10 and 11 represents the graphical representation of the testing performance of the DDM Level 2 classification model. In Fig. 11, DDM Level 2 classifier predicted Bacterial while the actual response was viral in row 1 and column 1.

Table 6

Comparative analysis with similar research

Authors	Class dataset	Technique	Image quantity	Recall	Precision	Accuracy
Vikash et al. [42]	Normal & Pneumonia	Transfer learning (various pre-trained models)	5232	99.5	93.28	96.39
Rajaraman [43]	Normal & Pneumonia	Custom VGG16 CNN model	5856	98.9	97.1	96.2
Toğaçar et al. [44]	Normal & Pneumonia	Deep CNN model	5849	96.83	96.88	96.84
Ayan et al. [45]	Normal & Pneumonia	VGG16 CNN model	5856	89.1	91.3	84.5
Saraiva et al. [46]	Normal & Pneumonia	Artificial neural network	5840	94.5	94.30	94.5
Khatri et al. [47]	Bacterial & Viral Pneumonia	EMD	144	89.5	80	83.3
Gu et al. [48]	Bacterial & Viral Pneumonia	Deep CNN Model	4882	77.55	88.86	80.4
Proposed DDM model (Level 1)	Normal & Pneumonia	Deep CNN Model	8800	97.9	97.55	97.45
Proposed DDM model (Level 2)	Viral & Bacterial	Custom DesnseNet201 model	6700	87.91	87.836	88.41

Figure 8.

Training and validation accuracy of level 1 classification (Normal or Pneumonia).

Figure 9.

Sample prediction of level 1 (Normal or Pneumonia) classifier.

Figure 10.

Training and validation accuracy of level 2 classification (Viral or Bacterial Pneumonia).

Figure 11.

Sample prediction of level 2 (Viral or Bacterial Pneumonia) classifier using customized DenseNet201.

Table 6 represents the performance of the proposed DDM Model with some of the similar and current work performed by the researchers. Rajaraman et al. [43] customized the VGG16 model and reported achieving an accuracy of 96.2% for the classification of Viral or Pneumonia. Ayan et al. [45] employed the VGG16 model to attain an accuracy of 84.5%, and it was outperformed by the customized VGG16 model introduced by Rajaraman et al. [43]. However, for the classification of Pneumonia Type (Viral or Bacterial), Gu et al. [48] proposed a Deep CNN based model to attain an accuracy of 80.4%, and it has three classes to choose from. In the same scenario, Khatri [46] employs the EMD model to outperform the CNN model introduced by Gu et al. [47] by achieving 83.3% accuracy. The suggested model was distributed in two levels to attain the maximum accuracy possible. Both groups did better than the other models by attaining 97.45% accuracy for Level 1 classification and 88.41% for Level 2 classification.

6. Conclusion

Many humans worldwide lose their lives every year due to Pneumonia, which is a potentially deadly disease if not treated as it should be and in the initial stage. Timely and accurate diagnoses in cognition with appropriate treatment may help save several lives. In third-world countries, where the health services are not up to the mark, there might be a lot of patients waiting in the outdoor emergencies for the appropriate diagnosis, and even then, it becomes a lot problematic to manage a huge amount of patients by doctors. In such cases, Computer-Aided Diagnosis (CAD) is a worthwhile option to speed up the diagnosed procedure. A Deep Neural Network model (DDM) is suggested in this effort to identify the Pneumonia type (Viral or bacterial). To attain this task, two levels were introduced in which Level 1 will classify the Normal or Pneumonia case using Chest X-ray image dataset, and if it classifies Pneumonia, the image information is transported to Level 2 for the classification of Viral or Bacterial Pneumonia. The quality parameters used to check its viabilities are Recall/Sensitivity, Precision, Accuracy and F1-Score. The performance of the suggested model for Level 1 classification was found to be improved than the current algorithm and attained the Accuracy, Recall, Precision and F1-Score of 97.41, 97.9, 97.55 97.731, respectively. For Level 2 classification, Accuracy, Recall, Precision and F1-Score are 88.41, 87.91, 87.836 and 88.242, respectively. Since regular DenseNet201 produced an excellent reply for the Chest X-ray image dataset, its custom model is utilized in this work for level 2 classification. The projected CAD-based DDM model will demonstrate a tool for radiologists to attain more image datasets and immediately diagnose Pneumonia with its type.

References

Bates

Campbell

Barton

McCracken

Morgan

Moses

, et al. Microbial Etiology of Acute Pneumonia in Hospitalized Patients. Chest. 1992 Apr; 101(4): 1005-12.

Jiang

Chen

, et al. Coronavirus Disease 2019 (COVID-19): A Perspective from China. Radiology. 2020 Feb 21; 296(2): 200490.

Omoumi

Ghani

Wong

Qiu

Liu

. The feasibility of utilizing the mid-energy in-line phase-contrast imaging system in the breast X-ray imaging. In: Chen

, editor. Biophotonics and Immune Responses XV. 2020 Mar 3, vol. 11241. pp. 1124114.

Sil Kar

Maity

. Gradation of diabetic retinopathy on reconstructed image using compressed sensing. IET Image Processing. 2018 Nov 1; 12(11): 1956-63.

Patil

Kuchanur

. Lung cancer classification using image processing. Int J Eng Innov Technol. 2012; 2(30): 37-42.

Lai

. Deep Learning for Medical Image Segmentation. arXiv:150502000 [cs]. 2015 May 8; Available from: https://arxiv.org/abs/1505.02000.

Esteva

Kuprel

Novoa

Swetter

Blau

, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017 Jan 25; 542(7639): 115-8.

Deng

Hinton

Kingsbury

. New types of deep neural network learning for speech recognition and related applications: an overview. In: 2013 IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE; 2013.

Demner-Fushman

Kohli

Rosenman

Shooshan

Rodriguez

Antani

, et al. Preparing a collection of radiology examinations for distribution and retrieval. J Am Med Inform Assoc. 2016; 23(2): 304-10. doi: 10.1093/jamia/ocv080.

10.

Tahir

Shahbaz Khan

Owais Tariq

. Performance analysis and comparison of faster R-CNN, mask R-CNN and ResNet50 for the detection and counting of vehicles. In: 2021 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS). IEEE; 2021.

11.

Abdelmaksoud

Shalaby

Mahmoud

Elmogy

Aboelfetouh

Abou El-Ghar

, et al. Precise identification of prostate cancer from DWI using transfer learning. Sensors. 2021; 21(11): 3664. doi: 10.3390/s21113664.

12.

Cernazanu-Glavan

Holban

. Segmentation of bone structure in X-ray images using convolutional neural network. Adv Electr Comput Eng. 2013; 13(1): 87-94.

13.

Chowdhury

MEH

Alzoubi

Khandakar

Khallifa

Abouhasera

Koubaa

, et al. Wearable real-time heart attack detection and warning system to reduce road accidents. Sensors. 2019; 19(12): 2780.

14.

Bhatt

Yadav

Sarvaiya

. Convolutional neural network based chest X-ray image classification for pneumonia diagnosis. In: Communications in Computer and Information Science. Singapore: Springer Singapore; 2020. pp. 254-66.

15.

Demner-Fushman

Kohli

Rosenman

Shooshan

Rodriguez

Antani

, et al. Preparing a collection of radiology examinations for distribution and retrieval. J Am Med Inform Assoc. 2016; 23(2): 304-10. doi: 10.1093/jamia/ocv080.

16.

Erdem

Aydin

. Detection of pneumonia with a novel CNN-based approach. Sakarya University Journal of Computer and Information Sciences. 2021; 26-34. doi: 10.35377/saucis.04.01.787030.

17.

Moujahid

Cherradi

Gannour

Bahatti

Terrada

Hamida

. Convolutional neural network based classification of patients with pneumonia using X-ray lung images. Adv Sci Technol Eng Syst J. 2020; 5(5): 167-75. doi: 10.25046/aj050522.

18.

Rahman

Chowdhury

MEH

Khandakar

Islam

Mahbub

, et al. Transfer learning with deep Convolutional Neural Network (CNN) for pneumonia detection using chest X-ray. Appl Sci. 2020; 10(9): 3233.

19.

Xie

Zhang

Cheng

. Predict pneumonia with chest X-ray images based on convolutional deep neural learning networks. J Intell Fuzzy Syst. 2020; 39(3): 2893-907.

20.

Zeng

Wang

Zhao

. Research on CNN-based models optimized by genetic algorithm and application in the diagnosis of pneumonia and COVID-19. bioRxiv. 2020; doi: 10.1101/2020.04.21.20072637.

21.

Çınar

Yıldırım

Eroğlu

. Classification of pneumonia cell images using improved ResNet50 model. Trait Du Signal. 2021; 38(1): 165-73.

22.

Liang

Zheng

. A transfer learning method with deep residual network for pediatric pneumonia diagnosis. Comput Methods Programs Biomed. 2020; 187(104964): 104964.

23.

Song

Han

X-Y

Montenegro-Marin

Krishnamoorthy

. Secure prediction and assessment of sports injuries using deep learning based convolutional neural network. J Ambient Intell Humaniz Comput. 2021; 12(3): 3399-410. doi: 10.1007/s12652-020-02560-4.

24.

Narin

Kaya

Pamuk

. Automatic detection of coronavirus disease (COVID-19) using X-ray images and deep convolutional neural networks. Pattern Anal Appl. 2021; 24(3): 1207-20. doi: 10.1007/s10044-021-00984-y.

25.

Zhang

Y-D

Satapathy

Guttery

Górriz

Wang

S-H

. Improved breast cancer classification through combining graph convolutional network and convolutional neural network. Inf Process Manag. 2021; 58(2): 102439. doi: 10.1016/j.ipm.2020.102439.

26.

Goel

Murugan

Mirjalili

Chakrabartty

. OptCoNet: an optimized convolutional neural network for an automatic diagnosis of COVID-19. Appl Intell. 2021; 51(3): 1351-66. doi: 10.1007/s10489-020-01904-z.

27.

Wang

Zhao

Wang

. Data-driven fault diagnosis method based on the conversion of erosion operation signals into images and convolutional neural network. Process Saf Environ Prot. 2021; 149: 591-601. doi: 10.1016/j.psep.2021.03.016.

28.

Kurata

Nishio

Moribata

Kido

Himoto

Otani

, et al. Automatic segmentation of uterine endometrial cancer on multi-sequence MRI using a convolutional neural network. Sci Rep. 2021; 11(1): 14440. doi: 10.1038/s41598-021-93792-7.

29.

Mohammed Sahan

Abbas

Abood

. A facial recognition using a combination of a novel one dimension deep CNN and LDA. Mater Today. 2021; doi: 10.1016/j.matpr.2021.07.325.

30.

ul Islam

. Using deep learning based methods to classify salt bodies in seismic images. J Appl Geophy. 2020; 178(104054): 104054. doi: 10.1016/j.jappgeo.2020.104054.

31.

Krizhevsky

Sutskever

Hinton

. ImageNet classification with deep convolutional neural networks. Commun ACM. 2017; 60(6): 84-90. doi: 10.1145/3065386.

32.

Xue

Han

. Identification, classification, and quantification of three physical mechanisms in oil-in-water emulsions using AlexNet with transfer learning. J Food Eng. 2021; 288(110220): 110220. doi: 10.1016/j.jfoodeng.2020.110220.

33.

Chen

Wan

Zhang

Chen

, et al. Medical image segmentation and reconstruction of prostate tumor based on 3D AlexNet. Comput Methods Programs Biomed. 2021; 200(105878): 105878. doi: 10.1016/j.cmpb.2020.105878.

34.

Han

. Detection and classification of marine mammal sounds using AlexNet with transfer learning. Ecol Inform. 2021; 62(101277): 101277. doi: 10.1016/j.ecoinf.2021.101277.

35.

Pan

Liang

Yang

. Classification of bacterial and viral childhood pneumonia using deep learning in chest radiography. In: Proceedings of the 3rd International Conference on Multimedia and Image Processing. New York, NY, USA: ACM; 2018.

36.

Allen-Zhu

. What can ResNet learn efficiently, going beyond kernels? arXiv [cs.LG]. 2019; Available from: http://arxiv.org/abs/1905.10337.

37.

Lin

Jegelka

. ResNet with one-neuron hidden layers is a Universal Approximator. arXiv [cs.LG]. 2018. Available from: http://arxiv.org/abs/1806.10909.

38.

Hammad

Pławiak

Wang

Acharya

. ResNet-Attention model for human authentication using ECG signals. Expert Syst. 2021; 38(6). doi: 10.1111/exsy.12547.

39.

Jiang

Chen

Zhang

Xiao

. Breast cancer histopathological image classification using convolutional neural networks with small SE-ResNet module. PLoS One. 2019; 14(3): e0214587. doi: 10.1371/journal.pone.0214587.

40.

LeCun

Kavukcuoglu

Farabet

. Convolutional networks and applications in vision. In: Proceedings of 2010 IEEE International Symposium on Circuits and Systems. IEEE; 2010. pp. 253-6.

41.

Jiang

Kot

. Deep coupled ResNet for low-resolution face recognition. IEEE Signal Process Lett. 2018; 25(4): 526-30. doi: 10.1109/lsp.2018.2810121.

42.

Chouhan

Singh

Khamparia

Gupta

Tiwari

Moreira

, et al. A novel transfer learning based approach for pneumonia detection in chest X-ray images. Appl Sci. 2020; 10(2): 559. doi: 10.3390/app10020559.

43.

Rajaraman

Candemir

Kim

Thoma

Antani

. Visualization and interpretation of convolutional neural network predictions in detecting pneumonia in pediatric chest radiographs. Appl Sci. 2018; 8(10): 1715. doi: 10.3390/app8101715.

44.

Toğaçar

Ergen

Cömert

Özyurt

. A deep feature learning model for pneumonia detection applying a combination of mRMR feature selection and machine learning models. IRBM. 2020; 41(4): 212-22. doi: 10.1016/j.irbm.2019.10.006.

45.

Ayan

Unver

. Diagnosis of pneumonia from chest X-ray images using deep learning. In: 2019 Scientific Meeting on Electrical-Electronics & Biomedical Engineering and Computer Science (EBBT). IEEE; 2019.

46.

Saraiva

Santos

Costa

Sousa

Ferreira

Valente

, et al. Models of learning to classify X-ray images for the detection of pneumonia using neural networks. In: Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies. SCITEPRESS – Science and Technology Publications; 2019.

47.

Khatri

Jain

Vashista

Mittal

Ranjan

Janardhanan

. Pneumonia identification in chest X-ray images using EMD. In: Lecture Notes in Networks and Systems. Singapore: Springer Singapore; 2020. pp. 87-98.

48.

Pan

Liang

Yang

49.

Kermany

. Labeled optical coherence tomography (OCT) and Chest X-ray images for classification. Mendeley. 2018. doi: 10.17632/rscbjbr9sj.2.

50.

Shen

Sun

. Squeeze-and-Excitation Networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE; 2018. pp. 7132-7141.

Dual deterministic model based on deep neural network for the classification of pneumonia

Abstract

Keywords

1. Introduction

2. Deep learning models

2.1 Convolutional Neural Network (CNN)

2.2.1 AlexNet

Table 4 Confusion matrix of AlexNet, ResNet18 and DDM along with their respective accuracies

References

Table 4
Confusion matrix of AlexNet, ResNet18 and DDM along with their respective accuracies