DKCNN: Improving deep kernel convolutional neural network-based COVID-19 identification from CT images of the chest

Abstract

BACKGROUND:

An efficient deep convolutional neural network (DeepCNN) is proposed in this article for the classification of Covid-19 disease.

OBJECTIVE:

A novel structure known as the Pointwise-Temporal-pointwise convolution unit is developed incorporated with the varying kernel-based depth wise temporal convolution before and after the pointwise convolution operations.

METHODS:

The outcome is optimized by the Slap Swarm algorithm (SSA). The proposed Deep CNN is composed of depth wise temporal convolution and end-to-end automatic detection of disease. First, the datasets SARS-COV-2 Ct-Scan Dataset and CT scan COVID Prediction dataset are preprocessed using the min-max approach and the features are extracted for further processing.

RESULTS:

The experimental analysis is conducted between the proposed and some state-of-art works and stated that the proposed work effectively classifies the disease than the other approaches.

CONCLUSION:

The proposed structural unit is used to design the deep CNN with the increasing kernel sizes. The classification process is improved by the inclusion of depth wise temporal convolutions along with the kernel variation. The computational complexity is reduced by the introduction of stride convolutions are used in the residual linkage among the adjacent structural units.

Keywords

Deep CNN temporal structural unit depth wise temporal pointwise

1 Introduction

Coronavirus (COVID-19) is a contagious disease that spread all over the world. It appeared in December 2019 in China millions of people were affected by this disease. The people infected by the virus have mild symptoms such as cough, cold, fever, and so on. Due to a severe attack, the respiratory system is affected and requires medical treatment. People are affected by cancer, diabetes, and other diseases, especially old age people who can be easily affected by COVID-19. The virus will spread through the mouth, nose while sneezing and by hands. It can be protected by washing hands, using sanitizer, mask, staying away from affected people, and being vaccinated at times [1, 2].

CT scan is usually used for imaging tests to identify the COVID-19. It is used for identifying many diseases and it is closely related to other diseases [3]. To identify the COVID-19 positive CT scan along with an imaging test is necessary. Image processing is the section of signal processing. It executes the performance of an image to do some actions. The image is data taken as an input and the required output will be the features of an image. CT is denoted as computed topography and it is also called a CAT scan. It initiates a 3-dimensional image and remodels numerous images that generate numerous images [4].

Machine learning algorithms are not programmed machines. It automatically collects authentic images from the earlier information. By using the algorithms, the machine language is converted to human-readable language [5]. It is hard to mark typical algorithms to monitor the task. It validates a huge amount of information. It can identify a set of images to recognize input data. The deep learning technique automatically selects and processes the images and the required output is determined. Artificial Intelligence is an instigator to achieve certain goals and design as per our own as humans. It is cyber espionage sustained by materials, as to resist by perception [6].

The machine and deep learning methods solve the complications of the following techniques which were demonstrated [7]. The applications of deep learning are it reduces image size and automatically detects output signals. Machine learning is a peculiar form of deep learning. In machine learning, the data are manually from the traditional data. But in deep learning, the data are automatically collected information from the computer. It requires a huge amount of data does not predict the precise data it is expensive and complexity is high [8]. Although the machine learning and deep learning methods offer automatic diagnosis of COVID-19 from the CT images, they face certain challenges. The major challenge faced by DL and ML algorithms is their complexity in training, particularly the need for large databases for accurate identification of COVID-19. Moreover, the conventional models face difficulty in capturing the intricate patterns and interconnections in the CT images, making it ineffective for real-time scenarios. Also, they require large computational resources and extensive databases for training; hence, implementing them is costly particularly in resource-constrained environments. Furthermore, conventional machine learning and deep learning models offer limited generalizability across diverse patient data, leading to inaccurate COVID-19 identification. In addition, they are less scalable and adaptable, making it less reliable for real-world scenarios.

To resolve these issues, we propose a novel Deep CNN-based SSA algorithm-based COVID-19 classification approach from the CT scan images. This proposed work combines the strengths of deep learning with kernel and optimization to enhance the accuracy in COVID-19 diagnosis. This innovative combination enables the system to learn hierarchical representations of features from CT images, allowing the system to capture complex patterns associated with COVID-19. In addition, the proposed Deep Kernel CNN effectively learns the relations within the data from a smaller database, reducing the computational burden, and resource requirements. The utilization of SSA optimization refines the training process of the classification model, reducing the computational efforts and minimizing computational time. Moreover, this combination enhances system’s generalization and reduces overfitting in COVID-19 detection. The key contributions of our proposed work are

•
The dataset related to the COVID-19 is collected and processed by a min-max preprocessing approach.
•
The preprocessed images are undergone the feature extraction step and extract the required features for the further process.
•
Then the proposed Deep CNN architecture is used to classify the images and to enhance the performances we utilize the SSA algorithm.

The article is arranged as below; in section 2 the relevant works are analyzed and listed their advantages and disadvantages along with it. The proposed framework is explained in section 3. The experimental setup and analysis are made in section 4. Finally, the work is summarized in section 5.
2 Literature review

Based on deep learning Zheng et al. [9] have proposed CT diagnosis to develop a precise analysis of coronavirus. CT scans are collected from covid positive cases, pneumonia cases, and no positive cases are analyzed for comparison purposes. A Details Relation Extraction neural network (DRENet) is designed to predict the images correctly. In the Feature Pyramid Network (FPN) the features are extricated from local and regional features. It helps to predict accurately the covid positive cases. However, additional image resources are needed to predict the diseases. Wang et al. [10] proposed an Image acquisition based authorized by Artificial Intelligence (IAAI) to identify COVID-19 cases. The model consists of three phases they are image acquisition, pre-scan preparation, and disease analysis. The formalities and instructions are given in the first stage. For analysis and reading the CT images are converted into images where the data can be easily determined. The images are efficient and safely determine the précised data. Thus, the image quality is to be improved. Nevertheless, the prediction of disease from low-quality images is a complicated task.

Dong et al. [11] have introduced the imaging characteristics to detect COVID-19. Positron emission tomography (PET) helps to evaluate consumptive diseases. The lung ultrasound allows for the initial harmless cases. It can monitor diseases and is comfortable for children and women. It is easily portable and the cost is less. The imaging characteristics are very fast and predict precise results. Meanwhile, it is difficult to predict the disease from a large dataset. Fan et al. [12] have described Infection Segmentation Deep Network (Inf-Net) method to evaluate the contaminated areas from CT slices. In this method, the contaminated areas are identified first and the shape of the local aspect is determined. The edge of boundaries is predicted and it is completely and easily enhanced. The model that detects insufficient data is to identify a correct infected part. However, the computational complexity and cost are a little bit high.

To develop a dual-sampling approach Ouyang et al. [13] proposed a 3D Convolutional Network (CNN) to identify COVID-19. To classify community-acquired pneumonia (CAP) as a dual-sampling approach is performed. For classification, CT images are given as input. This method predicts the dissimilarity of the contaminated regions. The COVID-19 is a contagious disease similar to community-acquired pneumonia that spreads relatively. This method is effectively performed in classification. The deformation in the classification should reduce. Roy et al. [14] have analyzed lung ultrasonography (LUS) by the deep learning techniques. The LUS images are used to evaluate and develop the segmentation methods. It predicts the LUS images and identifies the weak compulsive diseases. The region of interest has been utilized to classify the threshold from the input data. The segmentation method is performed for the pattern detection process. It also identifies asperity issues of regression. To determine the segments effectively groups of segments are evaluated. Anyhow, insufficiency of data is explicated for training data. Han et al. [15] have proposed an attention-based deep 3D multiple instance learning (AD3D-MIL) to identify the 3D chest CT weak labels. The data are divided into 3D cubes. Multiple instances can be explicated in one class. It reviews the characteristics of various findings. It efficiently optimizes in the 3D neural network to achieve the correct result. The 3D method is scalable and flexible in real-world chores. Thus, the sensitivity of this method should be enhanced.

3 Proposed framework

This section portrays the proposed COVID-19 classification from the CT scan images in an elaborated way. Figure 1 illustrates the schematic structure of the proposed framework in a clear manner.

Fig. 1

Schematic Structural framework of proposed work.

3.1 Data collection

The collection of CT scan images of COVID-19 affected cases is made from two publicly available datasets such as SARS-COV-2 Ct-Scan Dataset [16] and CT scan COVID Prediction [17]. The SARS-COV-2 Ct-Scan Dataset contains Covid-19 positive cases of 1257 scans and negative cases of 1232 scan images. The data are collected from the Sao Paul hospital, Brazil. CT scan COVID Prediction dataset is collected from the medRxiv, bioRxiv, NEJM, JAMA, Lancet, etc. The data are divided into training and testing images and include 502 training images and 257 testing images. The CT scan COVID Prediction dataset contains both Covid-19 and normal CT images comprising 502 Covid-19 and 257 normal images. The imbalance class in the database affects the training performance of the system. Hence, we utilized weighting techniques in the training phase to resolve these imbalance issues.

In this technique, we allocate higher weights to samples from the non-covid-19 instances (normal class) and lower weights to the majority class (Covid-19 class). The proposed methodology was validated using these two databases to demonstrate its reliability and generalization across diverse clinical data. We split the database in three different ratios like 90:10, 80:20, and 70:30 to assess the robustness of the model across different train-validation proportions. After splitting the database, we apply a data augmentation step to the training sequence for enhancing the model’s generalization ability. In this step, we apply different transformations including scaling, rotation, brightness, mixing images, cropping, etc., for assisting the model to learn invariant attributes in the samples. The sample images of SARS-COV-2 Ct-Scan Dataset [16] and CT scan COVID Prediction [17] are depicted in Figs. 2 and 3 respectively.

Fig. 2

Sample images from SARS-COV-2 Ct-Scan Dataset [16].

Fig. 3

Sample images from CT scan COVID Prediction [17].

3.2 Preprocessing

The gathered CT scan images are resized to the size of n × n and converted to the range of 0 to 1 by the normalization approach [18]. The normalization of training and testing data are normalized to increases the convergence speed of the neural network with the enhancement of stability of the method via an appropriate learning approach. Here we utilized the min-max normalization approach in order for mapping between 0 and 1 [19]. This can be defined as, $Y_{Norm} = \frac{Y - Y_{min}}{Y_{max} - Y_{min}}$ (1)

Here, Y_min and Y_max are the minimum and maximum values of the samples available in the CT scan images correspondingly.

3.3 Data extraction

In image processing, feature extraction is the process of capturing and extracting meaning information from the vast information of the input image. This step plays a significant role in highlighting the most relevant feature, which is essential for differentiating the COVID-19 and normal instances. In the proposed work, the major objective of using feature extraction is to ensure consistency and uniformity across all images in the preprocessed database. Firstly, image segmentation was performed for each CT image between the R-R periods, which maintains equal length between the R peaks. This step ensures uniformity across all images in the dataset by identifying the R-peak values based on neighboring pixel values and cropping the image’s edges. Secondly, zero padding is applied to set the extracted features into zero using zero padding [20]. The extracted features are set to be smaller or larger than the predetermined value. The steps involved in the feature extraction are explained below:

•
In this step, the R-peak value was detected based on the neighboring pixel values in the images. This identified R-peak value acts as the reference point for the image segmentation process. After R-peak detection, the images are segmented into smaller regions and each region indicates the most significant feature. $m_{min} = (u + v) / 2$ (2) $m_{max} = (v + w) / 2$ (3)

Here m denotes the extracted features, the sample image containing the adjacent peak values are indicated as u, v, w, and the peak value to be segmented is denoted as v. By detecting the R-peak value and analyzing the image’s neighboring pixels, we can extract the most significant features from the images for further analysis. The extracted feature sequence provides the most valuable information within the data, which is important for the classification model to differentiate normal and COVID-19 instances.
•
The segmented features are sometimes smaller or larger than the predetermined value and can be equalized for our proposed work. If the features are smaller than a padding process will be conducted and cropping is done if the value is greater than the predetermined value.

This step converts the raw images into meaningful representation, which helps the classification model to differentiate the normal and Covid cases. Also, it helps in reducing the noise and other unwanted attributes from the images; thereby reducing the overfitting challenge and enhances the generalization performance of the proposed model. Consequently, the segmentation of images provides better discrimination between the normal and Covid instances. In addition, by capturing and extracting the most salient attributes from the images reduces data dimension, and assists the classification model to focus on learning the most informative and valuable aspects of data, which increases the classification performances of the model.
3.4 Proposed Deep CNN for the diagnosis of COVID-19

After the completion of feature extraction, the next step is to classify the images based on Covid-19 positive and negative. For better classification, we utilized depth wise separable convolution architecture based on 1D structure and proposed a novel deep CNN based on that [21]. The following sections explain the proposed approach in a clear manner.

3.4.1 Depth wise separable convolution unit

Conventionally, the 2D CNN utilizes both spatial and inter-channel convolution are performed which might have led to the increase in the number of arithmetic operations and thus includes more kernels. This also increases the overfitting issues [22] along with computational cost. Meanwhile, in the depth wise separable 2D CNN both the operations are performed separately. The spatial convolution is carried out in each channel subsequently, the pointwise convolution [23] is made. These operations are made with inter-channel information. This type of operation is termed separable convolution.

In the depth wise separable unit, the temporal information is captured using the 1D temporal convolution. Meanwhile, the classification of different types of diseases is affected by overfitting issues. These can be overcome by the use of 1D temporal convolution. The comparison between the traditional and depth wise convolution is made by the following assumption.

Assume, the length of the input channel be h _i and the total number of the channel be N _i. Here, the kernel κ is used along with the 1D convolution to convert the input data (h _i, N _i) to (h _i, N _o). Here N _odenotes the output channel number. This consumes the computational cost of δ _s and the network parameters are given as ξ _s.

Here, $δ_{s} = h_{i} \times N_{i} \times N_{o} \times κ$ (4) $ξ_{s} = κ \times N_{i} \times N_{o}$ (5)

Here the computational cost is expressed as h _i × N _i × N _oalong with the network parameters of N _i × N _o. Then the total computational cost is given as, $δ_{d} = h_{i} \times N_{i} \times κ + N_{i} \times N_{o} \times h_{i}$ (6) $ξ_{d} = κ \times N_{i} + N_{i} \times N_{o}$ (7)

The output feature map of the convolution is altered with the aid of N _o/N _i factor. The mitigation of computational cost can be given as, $\frac{δ_{d}}{δ_{s}} = \frac{1}{N_{o}} + \frac{1}{κ}$ (8)

Meanwhile, the network parameter is mitigated to, $\frac{ξ_{d}}{ξ_{s}} = \frac{1}{N_{o}} + \frac{1}{κ}$ (9)

From this, it is noted that the depth wise separable convolution effectively mitigates the computational cost and the network parameters with the larger value of the kernel. Hence it can be used for our purposes.

3.4.2 Proposed unit

In our proposed structure the depth wise separable unit is designed to perform the temporal convolution and the inter-channel pointwise convolution. The proposed design is depicted to Fig. 4. The steps involved in the proposed convolution operation in the framework is elucidated below,

•
The inter-channel input CT image’s data are combined with the pointwise convolution operation and displayed in the space with the enhancement of a number of channels. Moreover, the optimization was selected precisely to enhance the depth factor.
•
Henceforth the convolution of separate temporal is performed along with the mapping of deeper features to capture each channel’s temporal information. The dimensionality of the kernel varies for each temporal convolution.
•
A transfer learning was introduced to improve the learning process of the developed mode, which leverages its knowledge encoded in pre-trained models for enhancing the feature extraction process. Here, we utilized the ResNet pre-trained model to extract the semantic features from the CT images. By assisting the Deep CNN learning process, the pre-trained model aims to improve the feature extraction process. Also, this incorporation enables the system to process different images obtained from diverse clinical settings.
•
The extracted feature incorporated with the temporal information of various channels is attained by pointwise convolution in the smaller area. Meanwhile, new features are obtained for each structural unit along with the conduction of temporal convolution and this can be acquired by the precise selection of reduction factor dimensionality.
•
The combination of input features along with the deeper structural unit can be obtained by the addition unit in the structure. The gradient issues can be overcome by the connection made between the input and output feature map.

Fig. 4
Proposed structural units with single temporal convolution.

Figure 5 illustrates the alterative structural unit for the depth wise separable unit. Here the operation of the single temporal convolution along with maximum kernel value is replaced with varying kernel [24] dimensions incorporated with the multiple temporal convolutions are conducted simultaneously utilizes the mapping of extended feature along with the convolution of pointwise. This will enable the combination of different temporal correlations gathered from the smaller to the higher time window values. This might have led to computational complexity. Subsequently, the values attained after the temporal convolutions are added prior to the pointwise convolution.

Fig. 5
The proposed structural unit with varying dimensionality.

While designing the classification network of deep CNN-based SSA architecture we have to consider the varying structural units of depth wise convolution. Here we proposed a multiple parallel temporal kernel-based approach which is discussed below.
3.5 Proposed Deep CNN based SSA architecture

The proposed architecture involves convolution of multiple temporal operations that are parallel conducted utilizing the different sizes of the kernel at an instance. A stride temporal convolution is conducted to mitigate the computational complexity along with this; it also mitigates the mapping of output feature-length. The steps involved in this process is explained below:

•
The process begins with the pointwise convolution to increase the factor by 2 and then the nonlinear activation function which normalizes the input used.
•
Henceforth, the normalized data are forwarded to the four parallel paths to conduct the temporal convolution which utilizes the strides of 2. As mentioned above this step mitigates the computational complexities and provides better features for further actions.
•
The output features maps obtained from the four parallel paths from the nonlinear activation and normalization function are added to provide an increasing number of channels.
•
Then the added features are undergone pointwise convolution and are placed in the smaller area. Then the output feature maps are enhanced to 32 which is higher than the input feature map.
•
Subsequently, the convolution of stride pointwise is conducted to produce the finalized mapping of the output feature by combining the output of the pointwise convolution and input feature map. The convolutional layers in the developed model aim to capture the most relevant attributes from the CT images by analyzing its spatial and hierarchical patterns and correlations.
•
By performing convolution operation across the input images enables the system to automatically learn and understand the local patterns and structures, which are significant for differentiating the normal and Covid-19 cases. Generally, these convolutional layers act like a filter, which slides over the CT images and captures the meaningful features like textures, structures, edges, etc. This convolution operation provides the hierarchical feature representation, which assists the model in learning and understanding the variations and differences between the normal and diseased images. Figure 6 illustrates the proposed architecture and elucidates the process in detail. At first, the input is fed into the standard convolution block and forwarded to the Deep CNN unit block. This process will increase the feature map and thereby increase the depth of the network. Thus the depth increases to 160 channels and the length mitigates to 8. Our proposed approach effectively extracts the features with multiple temporal kernels depicted as Fig. 7. The choice of the kernel directly influences the performances of the developed model. Since, the kernels play a crucial role in capturing the most intricate and temporal interconnections within the data, the type of kernel directly hits the classification performance of the model. In addition to this, the kernel size is also an influencing factor in capturing the characteristics within the images. The kernel with larger size captures broader temporal information, while the smaller sized kernels extract the finer information. In the proposed work, we designed the Deep CNN with increasing kernel sizes for capturing a wide range of temporal attributes in the CT images, which enhances the system’s capacity to distinguish Covid-19 and normal cases. Also, it can be used to combine the features that are obtained from observation windows. Moreover, it is also used to snap the complex functionality of the information to differentiate the features of different Covid-19 classes.

Fig. 6
Proposed architecture using multiple temporal structural units.

Fig. 7
Proposed DeepCNN.

Meanwhile, the convergences [26] of the proposed architecture are improved by using the activation function optimizer which is explained below. The nonlinearity between the convolution layers and the exponential linear unit (ELU) [25] is introduced to analyze the COVID-19. It can be given as: $ELU : h (a, y) = {\begin{matrix} a (e^{y} - 1), if y ⩽ 0 \\ y, if y > 0 \end{matrix}$ (10)

The fast convergence is obtained by setting a = 0.2. The SoftMax layer is utilized to detect the Covid-19 cases. Then the performances of our proposed work are enhanced by the adoption novel Salp Swarm algorithm. The proposed DeepCNN based SSA can be used for the effective classification of COVID-19 and normal cases.

Salp swarm algorithm (SSA):

The new metaheuristic algorithm is named SSA [27, 32, 33]. It is a bio-inspired optimizer to solve optimization problems and it is proposed by Mirjalili et al. In the deep sea it navigates, forages and this algorithm is formulated by the thronging behavior of the salps. This algorithm is also called a salps chain. The salps chain is classified into two, the first one is the leader and the second is the follower. Firstly, the leader guides the entire chain. Secondly, the followers are allowed to follow the entire chain by one another. The n-dimensional position of the leader in search context is defined as: $y_{j}^{1} = {\begin{matrix} S_{1} + d_{1} ((u_{j} - l_{j}) d_{2} + l_{j}), d_{3} ⩾ 1 \\ S_{1} - d_{1} ((u_{j} - l_{j}) d_{2} + l_{j}), d_{3} < 1 \end{matrix}$ (11)

Here, the first position of salp is denoted as $y_{j}^{1}$ and the leader in the jth position. The food source position is S ₁. The j^th positions of lower and upper bound u _j and l _j respectively and the random numbers are given as, d ₁, d ₂, d ₃.

The exploitation and exploration of coefficient d ₁ for better search is expressed as, $d_{1} = 2 b^{(- 4 i / N)^{2}}$ (12)

The current and maximum number of iterations is given as i and Nrespectively. The random numbers range from 0 and 1 of coefficients d ₂, d ₃. The follower’s position is expressed as follows, $y_{k}^{j} = 1 / 2 {Bt}_{2} + u_{0} t, j ⩾ 2 z$ (13)

The position of the followers is given as $y_{k}^{j}$ . $u_{0}$ is denoted as the initial speed and time t. B variable is expressed as: $B = u_{final} / u_{0}$ (14) $u = y - y_{0} / t$ (15)

When $u_{0}$ = 0. The equation is denoted as: $y_{k}^{j} = 1 / 2 (y_{k}^{j} + y_{k}^{j - 1})$ (16)

The salp swarm algorithm reduces the computational efforts and solves multi-scale problems. Thus the proposed work can be effectively used to classify the CT scan images for the diagnosis of Covid-19 diseases.
4 Experimental analysis

This section elaborately presents the experimental analysis of our proposed work. It includes the experimental setup, performance analysis, and comparative analysis.

4.1 Experimental setup

For the experimental purpose, we have utilized a system with the model of i5 7300HQ CPU@2.50 GHz, GTX 1650 4GB graphics system. For implementation purposes, we have taken the Keras library from Python with the Tensor flow in the Anaconda environment used as backend operations [31].

4.2 Evaluation metrics

To analyze the effectiveness of the normally classified Covid-19 disease images and normal images we have taken the evaluation metrics such as sensitivity, specificity, accuracy, Matthews Correlation Coefficients, and precision. Since we deployed the proposed model for binary classification, the evaluation of these metrics provides a clear performance of the model in differentiating the normal and COVID-19 instances. These performance metrics are assessed by estimating how well the classification model discriminates between the COVID-19 cases and normal. These performance metrics are determined through the intensive assessment of true positive, true negative, false positive and false negative parameters. The true positive indicates how well the proposed model identifies the COVID-19 instances, while true negative measures the model’s capacity to correctly identify the normal cases. On the other hand, the false positive represents the scenario where the developed algorithm incorrectly identifies the COVID-19 instance as normal and false negative defines the scenario where the model incorrectly detects the normal case as COVID-19. The definition for these metrics areelucidated below,

•Accuracy (A)

Accuracy can be defined as how exactly the CT images are classified as Covid-19 and normal. It quantifies how accurately the developed algorithm identifies the true positive and true negative instances from the total instances. It can be given as, $A = \frac{XX + XY}{XX + XY + ZX + ZY}$

•Sensitivity (SN)

Sensitivity is defined as the accurately classified Covid-19 images from the CT image datasets to the totally available Covid-19 images. This measures the model’s capacity to correctly identify the COVID-19 instances (true positive) from the total positive cases. It can be given as, $SN = \frac{XX}{XX + XY}$

•Specificity (SP)

Specificity is defined as the accurately classified normal image from the totally available normal images in the CT scan datasets. It defines the model’s capacity to identify the true negative instances (normal) correctly from the total true instances. It can be evaluated as, $SN = \frac{XY}{XX + XY}$

•Precision

Precision is defined as the probability of classifying the Covid-19 images from the actually available Covid-19 datasets. Precision measures the ratio of true positive predictions among all positive predictions made by the model. It can be given as, $Pr ecision = \frac{XX}{XX + ZX}$

•Matthews Correlation Coefficients (MCC)

MCC is determined as the difference between actually classified images and the expected classification accuracy. This metrics provides the balanced performance of the developed model in COVID-19 classification considering both true and false cases. It can be given as, $MCC = \frac{XX . XY + ZX . ZY}{\sqrt{(XX + ZY) . (XX + ZX) . (ZY + ZX) . (XY + ZY)}}$

Here, ZY denotes the false-negative value, ZX denotes the false positive values, XX denotes the true positive values, XY denotes the true negative. By analyzing these parameters, we can determine the model’s capacity to differentiate normal and COVID-19 instances.

4.3 Network configuration

The network configuration of the proposed deep CNN based SSA approach is listed in Table 1. The performance accuracy of different CNN architectures based on the training and validation is illustrated in Tables 2 and 3. Table 2 illustrates the variation of performance based on various CNN architectures. Table.3 illustrates the performance analysis based on the train-validation splits. The goal of the developed model is to categorize the image as normal or covid-19. Firstly, the dataset containing both normal and covid-19 images are prepared for training the model. Data preparation involves steps like preprocessing, labeling, and splitting. Further, the developed classifier model is trained using the training subset to understand the pattern and correlation between normal and covid-19 images. Then, the trained model was evaluated using the metrics like accuracy, precision, recall and f-measure. The testing of the model over unseen images provides model’s generalization ability. Here, we compared the performance of the proposed technique with different CNN models to validate its efficiency in covid-19 classification. The performances are evaluated under different validation sets. To analyze the performance, we took the CNN architecture such as ResNet [28], EfficientNet [29], and DesNet [30].

Table 1
Network configuration

Configuration of network

U V W X

Input CT scan images

Front-end DeepCNN model with fine tuned

Back-end with four various kernels

Convol-512-3-1 Convol -512-3-2 Convol -512-3-3 Convol -512-3-4

Convol-512-3-1 Convol -512-3-2 Convol -512-3-3 Convol -512-3-4

Convol-512-3-1 Convol -512-3-2 Convol -512-3-3 Convol -512-3-4

Convol -256-3-1 Convol -256-3-2 Convol -256-3-3 Convol -256-3-4

Convol -128-3-1 Convol -128-3-1 Convol -128-3-3 Convol -128-3-4

Convol -64-3-1 Convol -64-3-2 Convol -64-3-3 Convol -64-3-4

Mean pooling globally

FC-256

FC-256

∑

Configuration of network
U	V	W	X
Input CT scan images
Front-end DeepCNN model with fine tuned
Back-end with four various kernels
Convol-512-3-1	Convol -512-3-2	Convol -512-3-3	Convol -512-3-4
Convol-512-3-1	Convol -512-3-2	Convol -512-3-3	Convol -512-3-4
Convol-512-3-1	Convol -512-3-2	Convol -512-3-3	Convol -512-3-4
Convol -256-3-1	Convol -256-3-2	Convol -256-3-3	Convol -256-3-4
Convol -128-3-1	Convol -128-3-1	Convol -128-3-3	Convol -128-3-4
Convol -64-3-1	Convol -64-3-2	Convol -64-3-3	Convol -64-3-4
Mean pooling globally
FC-256
FC-256
∑

Table 2

Performance analysis of various CNN architectures based on the training and validation accuracy

Various CNN architecture	Accuracy (%)-training	Accuracy (%)-validation
ResNet	96.25	87.45
EfficientNet	97.24	89.76
DesNet	95.55	86.21
Proposed	98.30	95.20

Table 3

Performance analysis of various train and testing dataset splitting based on the training and validation accuracy

Splitting of train-validation	Accuracy (%)-training	Accuracy (%)-validation
90–10	99.67	92.01
80–20	98.01	90.45
70–30	99.22	88.56

4.4 Performance analysis based on the Network configuration

The performance evaluation based on the network configuration for the different CNN architecture for the dataset Ct-Scan Dataset [16] is illustrated in Table 4. From the table, it is observed that the proposed method achieves better accuracy, precision, MCC than the others with the values of 98.30%, 98.04%, and 99.47% respectively.

Table 4
Performance analysis based on the network configuration for Ct-Scan Dataset [16]

Various CNN architecture Accuracy (%) Precision (%) MCC (%)

ResNet 78.34 84.78 89.56

EfficientNet 84.27 78.90 90.38

DesNet 92.09 94.27 95.67

Proposed 98.30 98.04 99.47

Various CNN architecture	Accuracy (%)	Precision (%)	MCC (%)
ResNet	78.34	84.78	89.56
EfficientNet	84.27	78.90	90.38
DesNet	92.09	94.27	95.67
Proposed	98.30	98.04	99.47

Meanwhile, the performance analysis based on the network configuration for the CT Scan COVID Prediction [17] is illustrated in Table 5. Here we have taken the CNN architecture such as ResNet, EfficientNet, DesNet, and proposed approaches. It is observed that the proposed approach accomplishes accuracy, precision, and MCC such as 98.90%, 99.03%, and 99.21% correspondingly.

Table 5

Performance analysis based on the network configuration for CT Scan COVID Prediction [17]

Various CNN architecture	Accuracy (%)	Precision (%)	MCC (%)
ResNet	79.08	87.56	93.63
EfficientNet	89.05	90.89	96.41
DesNet	94.89	95.87	95.49
Proposed	98.90	99.03	99.21

4.5 Comparative study in terms of evaluation metrics

For the comparative study, we have taken the state-of-art works such as DRENet, IAAI, Inf-Net, AD3D-MIL and compared them with our proposed work. The comparative analysis based on the accuracy, sensitivity, specificity, precision, and MCC is listed in Table 6. From the table, the technique DRENet attains the sensitivity, accuracy, precision, specificity, and MCC of 92.56%, 88.30%, 97.56%, 84.67%, and 86.45% respectively. On the other hand, our proposed approach attains the maximum evaluation values of equal to 98.38%, 99.12%, 98.56%, 99.65%, and 97.99% of accuracy, sensitivity, specificity, precision, and MCC correspondingly. From all values, our proposed is the optimal solution due to the selection of multiple temporal based on the varying kernel-based CNN approach, and the parameters are tuned by using the adopted SSA approach. The involvement of SSA also reduces the computational iterations and thus the complexity is also reduced to a great extent.

Table 6
Comparative study in terms of evaluation metrics

Techniques Accuracy (%) Sensitivity (%) Specificity (%) Precision (%) MCC(%)

DRENet 88.30 92.56 84.67 97.56 86.45

IAAI 83.27 94.90 86.66 94.34 90.78

Inf-Net 93.29 86.30 89.09 96.45 94.33

AD3D-MIL 89.12 93.29 94.29 92.71 91.20

Proposed 98.38 99.12 98.56 99.65 97.99

Techniques	Accuracy (%)	Sensitivity (%)	Specificity (%)	Precision (%)	MCC(%)
DRENet	88.30	92.56	84.67	97.56	86.45
IAAI	83.27	94.90	86.66	94.34	90.78
Inf-Net	93.29	86.30	89.09	96.45	94.33
AD3D-MIL	89.12	93.29	94.29	92.71	91.20
Proposed	98.38	99.12	98.56	99.65	97.99

4.6 Discussion

In this study, we proposed a hybrid classification model leveraging the advantages of deep learning, meta-heuristic optimization and kernel methods. The primary objective of this study is to accurately identify and classify normal and COVID-19 instances from the CT images with less computational time. Initially, the CT images containing both normal and Covid-19 instances are collected and fed into the system. The images undergo preprocessing steps, which improves the image quality by reducing the noises, and other backgrounds. Further, feature extraction was performed to extract the most significant attributes from the preprocessed images, which enables the classification model to focus on most informative and valuable attributes for classification. The classification model integrates the Deep CNN with the kernel methods for learning the patterns, and interconnections within the data for differentiating normal and COVID-19 cases. This unique combination enables the system to effectively learn the complex intricate patterns, and hierarchical representations within the CT images, leading to accurate identification of Covid-19. Furthermore, the SSA optimization was utilized to improve the training of the classification model by continuously refining its parameters to its finest value. This optimization process enables the system to achieve better convergence and reduces the computational time, improving the overall classification performances.

The presented study was trained and validated using the publicly available SARS-COV-2 Ct-Scan Dataset and CT scan COVID Prediction databases. The experimental analysis depicts that the designed algorithm achieved improved performances like high accuracy, improved precision, etc. Further, we made a comparative study with the conventional models, and it validated that the developed framework obtained better results than existing techniques. These improved performances of the developed algorithm make it effective and reliable for real-time COVID-19 classification in clinical settings. However, the utilization of artificial intelligence in the diagnosis of life-threatening diseases creates trust issues. Therefore, ensuring interpretability and explainability is significant for accurate and reliable decision-making in clinical fields. Hence, the future study should be done on improving the interpretability and explainability by introducing visualization techniques.

5 Conclusion

The work in this article is based on the DeepCNN based SSA approach to detect the Covid-19 and normal images from the CT scan image datasets. The proposed Deep CNN architecture is composed of a structural unit known as Pointwise-Temporal-Pointwise Convolution. This approach uses depth wise separable convolution and at the first stage, the collected datasets are preprocessed to remove the unwanted noises and blurred data. Subsequently, the features are extracted and fed into the proposed Deep CNN architecture. Meanwhile, the feature mapping is mitigated by the usage of the different temporal windows in parallel. Thus, the classifications of CT scan images for the diagnosis of Covid-19 are effectively possible and attain higher accuracy. Further, the experiments are conducted in Python using the Ct-Scan Dataset, and CT scan COVID Prediction dataset. For comparative analysis, we have taken different state-of-art works such as DRENet, IAAI, Inf-Net, and AD3D-MIL. From the performance analysis, it is concluded that the proposed work achieves better evaluation measurements and therein effectively reduces the computational complexity.

References

World Health Organization, The Corona Virus Disease (COVID-19), (2019).

Sun

Dai

Deng

Huang

Sun

Prevalence and severity of corona virus disease (COVID-19): A systematic review and meta-analysis, Journal of Clinical Virology 127 (2020), pp. 104371.

Zhao

Zhang

Xie

Covid-ct-dataset: a ct scan dataset about covid-19. arXiv preprint arXiv:2003.13865490 (2020).

Afshar

Heidarian

Enshaei

Naderkhani

Rafiee

M.J.

Oikonomou

Fard

F.B.

Samimi

Plataniotis

K.N.

Mohammadi

COVID-CT-MD, COVID-19 computed tomography scan dataset applicable in machine learning and deep learning, Scientific Data 8(1) (2021), pp. 1–8.

Brinati

Campagner

Ferrari

Locatelli

Banfi

Cabitza

Detection of COVID-19 infection from routine blood exams with machine learning: a feasibility study, Journal of Medical Systems 44(8) (2020), pp. 1–12.

Alimadadi

Aryal

Manandhar

Munroe

P.B.

Joe

Cheng

Artificial intelligence and machine learning to fight COVID-19, Physiological genomics 52(4) (2020), pp. 200–202.

Panwar

Gupta

P.K.

Siddiqui

M.K.

Morales-Menendez

Singh

Application of deep learning for fast detection of COVID-19 in X-Rays using nCOVnet, Chaos, Solitons & Fractals 138 (2020), 109944.

Zhao

Jiang

Qiu

Deep learning for COVID-19 detection based on CT images, Scientific Reports 11(1) (2021), pp. 1–12.

Song

Zheng

Zhang

Huang

Chen

Wang

Zhao

Chong

Shen

Deep learning enables the accurate diagnosis of novel coronavirus (COVID-19) with CT images, IEEE/ACM Transactions on Computational Biology and Bioinformatics 18(6) (2021), pp. 2775–2780.

10.

Shi

Wang

Shi

Wang

Tang

Shi

Shen

Review of artificial intelligence techniques in imaging data acquisition, segmentation, and diagnosis for COVID-19, IEEE Reviews in Biomedical Engineering 14 (2020), pp. 4–15.

11.

Dong

Tang

Wang

Hui

Gong

Xue

Liao

Chen

Yang

Jin

The role of imaging in the detection and management of COVID-19: a review, IEEE Reviews in Biomedical Engineering 14 (2020), pp. 16–29.

12.

Fan

D.P.

Zhou

G.P.

Zhou

Chen

Shen

Shao

Inf-net: Automatic covid-19 lung infection segmentation from ct images, IEEE Transactions on Medical Imaging 39(8) (2020), pp. 2626–2637.

13.

Ouyang

Huo

Xia

Shan

Liu

Yan

Ding

Yang

Song

Shi

Dual-sampling attention network for diagnosis of COVID-19 from community-acquired pneumonia, IEEE Transactions on Medical Imaging 39(8) (2020), pp. 2595–2605.

14.

Roy

Menapace

Oei

Luijten

Fini

Saltori

Huijben

Chennakeshava

Mento

Sentelli

Peschiera

Deep learning for classification and localization of COVID-19 markers in point-of-care lung ultrasound, IEEE Transactions on Medical Imaging 39(8) (2020), pp. 2676–2687.

15.

Han

Wei

Hong

Cong

Zhu

Wei

Zhang

Accurate screening of COVID-19 using attention-based deep 3D multiple instance learning, IEEE Transactions on Medical Imaging 39(8) (2020), pp. 2584–2594.

16.

https://www.kaggle.com/plameneduardo/sarscov2-ctscan-dataset.

17.

https://www.kaggle.com/c/covidct/data.

18.

Liu

Tang

AdaDM: Enabling Normalization for Image Super-Resolution, arXiv preprint arXiv:2111.13905 2021.

19.

Hoang

N.-D.

Detection of surface crack in building structures using image processing technique with an improved Otsu method for image thresholding, Advances in Civil Engineering 2018 (2018).

20.

Nixon

Aguado

, Feature extraction and image processing for computer vision Academic press. 2019.

21.

Mahmud

Fattah

S.A.

Saquib

Deeparrnet: An efficient deep cnn architecture for automatic arrhythmia detection and classification from denoised ecg beats, IEEE Access 8 (2020), pp. 104788–104800.

22.

Ying

An overview of overfitting and its solutions. In Journal of Physics: Conference Series, 1168(2) (2019), pp. 022022. IOP Publishing.

23.

Hua

B.S.

Tran

M.K.

Yeung

S.K.

Pointwise convolutional neural networks. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018), (pp. 984–993).

24.

Bobodzhanov

A.A.

Safonov

V.F.

Regularized asymptotic of solutions to integrodifferential partial differential equations with rapidly varying kernels, Ufa Math. J 10(2) (2018), 3–13.

25.

Panda

Naskar

Pal

Exponential linear unit dilated residual network for digital image denoising, Journal of Electronic Imaging 27(5) (2018), 053024.

26.

Wang

Liu

Y.J.

Huang

Ren

Finite-time convergence adaptive neural network control for nonlinear servo systems, IEEE Transactions on Cybernetics 50(6) (2019), pp. 2568–2579.

27.

Faris

Mirjalili

Aljarah

Mafarja

Heidari

A.A.

Salp swarm algorithm: theory, literature review, and application in extreme learning machines, Nature-Inspired Optimizers (2020), pp. 185–199.

28.

Leelaruji

Teerakawanich

, March. Short Term Prediction of Solar Irradiance Fluctuation Using Image Processing with ResNet. In 2020 8th International Electrical Engineering Congress (iEECON) (2020), (pp. 1–4). IEEE.

29.

Kumar

Abhishek

Singh

COVID-19 Detection from Chest X-Rays and CT scans using Dilated Convolutional Neural Networks. In 2021 10th IEEE International Conference on Communication Systems and Network Technologies (CSNT) (2021), (pp. 369–374). IEEE.

30.

Yuan

Chen

Realtime CNN-based Keypoint Detector with Sobel Filter and CNN-based Descriptor Trained with Keypoint Candidates. arXiv preprint arXiv:2011.02119 2020.

31.

Kumar

Abhishek

Singh

COVID-19 Detection from Chest X-Rays and CT Scans using Dilated Convolutional Neural Networks, 2021 10th IEEE International Conference on Communication Systems and Network Technologies (CSNT). IEEE 2021.

32.

Albert

J.R.

et al., A Symmetric Solar Photovoltaic Inverter to Improve Power Quality Using Digital Pulsewidth Modulation Approach, Wireless Pers Commun 130 (2023), 2097–2097. https://doi.org/10.1007/s11277-023-10372-w.

33.

Murugesan

et al., A Hybrid Deep Learning Model for Effective Segmentation and Classification of Lung Nodules from CT Images, Journal of Intelligent and Fuzzy System 42(3) (2021) pp. 2667–26791. 10.3233/JIFS-212189.

DKCNN: Improving deep kernel convolutional neural network-based COVID-19 identification from CT images of the chest

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSION:

Keywords

1 Introduction

3 Proposed framework

3.4.1 Depth wise separable convolution unit

4.1 Experimental setup

4.2 Evaluation metrics

4.3 Network configuration

Table 4 Performance analysis based on the network configuration for Ct-Scan Dataset [16] Various CNN architecture Accuracy (%) Precision (%) MCC (%) ResNet 78.34 84.78 89.56 EfficientNet 84.27 78.90 90.38 DesNet 92.09 94.27 95.67 Proposed 98.30 98.04 99.47

5 Conclusion

References

Table 4
Performance analysis based on the network configuration for Ct-Scan Dataset [16]

Various CNN architecture Accuracy (%) Precision (%) MCC (%)

ResNet 78.34 84.78 89.56

EfficientNet 84.27 78.90 90.38

DesNet 92.09 94.27 95.67

Proposed 98.30 98.04 99.47