key: cord-0078576-7uescwik authors: Liu, Bing; Nie, Xuan; Li, Zhongxian; Yang, Shihong; Tian, Yushu title: Evolving deep convolutional neural networks by IP-based marine predator algorithm for COVID-19 diagnosis using chest CT scans date: 2022-05-24 journal: J Ambient Intell Humaniz Comput DOI: 10.1007/s12652-022-03901-1 sha: bd1cd4a9399b982b38aa1b4141593ecea74b2441 doc_id: 78576 cord_uid: 7uescwik This paper proposes an optimal structured deep convolutional neural network (DCNN) based on the marine predator algorithm (MPA) to construct a novel automatic diagnosis platform that may help radiologists identify COVID-19 and non-COVID-19 patients based on CT scan categorization and analysis. The goal is met with the help of three modifications based on the regular MPA. First, a novel encoding scheme based on Internet Protocol (IP) addresses is proposed, followed by introducing an Enfeebled layer to build a variable-length DCNN. Finally, the learning process divides big datasets into smaller chunks that are randomly evaluated. The proposed model is compared to the COVID-CT and SARS-CoV-2 datasets to undertake a complete evaluation. Following that, the performance of the developed model (DCNN-IPMPA) is compared to that of a typical DCNN and seven variable-length models using five well-known comparison metrics, as well as the receiver operating characteristic and precision-recall curves. The results show that the DCNN-IPMPA outperforms other benchmarks, with a final accuracy of 97.21% on the SARS-CoV-2 dataset and 97.94% on the COVID-CT dataset. Also, timing analysis indicates that the DCNN processing time is the best among all benchmarks as expected; however, DCNN-IPMPA represents a competitive result compared to the standard DCNN. Due to the widespread of COVID-19 and its variations in all nations, the timely and exact diagnosis of COVID-19 is essential. The initial stage in the COVID-19 diagnosis is to take the reverse transcription-polymerase chain reaction (RT-PCR) (Hu et al., 2021a) ; however, in a few situations, the RT-PCR test of a subject is negative, yet the associated patient obtains difficulty in breathing (Wu et al., 2021a, b) . The chest CT scan is advised as a second-line diagnostic in this situation. If the lungs are heavily infected with COVID-19, its diagnosis by thorough observation can be easily made. However, when the disease is in its early stages, a visual diagnosis of this condition is accompanied by some doubt. COVID-19 lung damage has recently been identified using X-rays and CT scans . However, the diagnostic accuracy is dependent on the expert's assessment (Zhang et al., 2022) . Since changes in pixels' value may be better identified by image processing techniques as a quantitive method, this research aims to automatically and more accurately identify the contaminated CT images (Cao et al., 2022; Liu et al., 2022a, b) . Deep learning (DL) algorithms can automatically identify this condition to address the mentioned problem (Saffari et al., 2020) . CT Scans reveal the bones, blood vessels, and other internal organs. As a result, they enable clinicians to assess the size and structure of internal organs. Compared to X-rays, CT Scans may isolate a specific body location without damaging nearby areas, and they can also reveal detailed images of the patient's body (Shah et al., 2021) . This data can be evaluated to determine the exact location of the problem. As a result, deep learning techniques have made extensive use of the potential of chest CT scans for early detection and diagnosis of COVID-19 (Afshar et al., 2021; Dai et al., 2020; Shah et al., 2021) . Grewal et al., for example, used DenseNet and RNNs to interpret brain CT images (Grewal et al., 2018) . Song et al. developed three deep neural models for lung cancer detection (SAE, DNN, and CNN) (Song et al., 2017) . They discovered that the DCNN architecture surpasses the other models regarding accuracy. Gonzales et al. employed deep learning to investigate the neural system's ability to identify lung diseases and predict acute respiratory problems (González et al., 2018) . During the pandemic, it was discovered that CT Scanbased analysis methods accurately diagnosed COVID-19 disease. Zhao et al., for example, examined the association between chest CT scan data and pneumonia in COVID-19 patients . Due to the scarcity of acceptable COVID-19 CT Scan datasets, Zhao et al. constructed a freely accessible dataset consisting of 363 healthy persons and 349 COVID-19 occurrences from 216 patients (Jiang et al., 2019) . Gozes et al. investigated the detection and tracking of COVID-19-positive cases. The accuracy of this improved model has been evaluated to be 95% (Gozes et al., 2020) . Zheng et al., on the other hand, proposed a deep learning system for COVID-19 detection in 3D CT images (Zheng et al., 2020) . Ai et al. demonstrated 97 percent sensitivity when using a pre-trained UNet to predict infection in 630 patients' CT scans (Ai et al., 2020) . COVID-CT was perhaps the first freely released dataset of this type. Following the presentation, this data set was used by a variety of researchers (Breban et al., 2013; Liu et al., 2020; Ozturk et al., 2020) . Previously, the best F1-score, accuracy, and AUC were 86.5%, 85%, and 94% (Breban et al., 2013) . Soares et al. (Angelov and Almeida Soares, 2020) presented a dataset of 2482 SARS-CoV-2 CT scans. Nonetheless, the aforementioned techniques need considerable processing time . As a result, a precise and rapid COVID-19 detector is necessary. Following a comprehensive investigation of the methods recently published for identifying and diagnosing COVID-19, it is clear that the DCNN is one of the most widely used techniques (Liu et al., 2017) ; thus, this research recommends utilizing DCNN's extraordinary potential as a detector of COVID-19 scans. DCNN design complexity has been reduced by metaheuristics (Khishe and Mosavi, 2020a, b; Qiao et al., 2021) , which develop an architecture without the assistance of a human designer (Rastogi and Choudhary, 2019) . Numerous metaheuristics have been effectively studied and applied, including genetic programming (GP) (Suganuma et al., 2017) , particle swarm optimization (PSO) Wang et al., 2018; J. Wu et al., 2021a, b) , Position-transitional particle swarm optimization , sine-cosine algorithm (SCA) C. Wu et al., 2021a, b) , chimp optimization algorithm (ChOA) (Hu et al., 2021b) , salp swarm algorithm (Khishe and Mohammadi, 2019) , and genetic algorithms (GAs) (Liu et al., 2022a, b) . However, learning from large data sets is impracticable due to the high cost of computing and the time of the learning process . Perhaps, the EvoCNN was the first model that employed metaheuristic algorithms to evolve DCNN structures (Stanley and Miikkulainen, 2002) . The EvoCNN helps save time by learning the classifier ten epochs rather than 25,600 epochs like the LEIC does. Regardless of the advantages of different metaheuristic algorithms, the No-Free-Lunch (NFL) theorem (Webb et al., 2011) states that no metaheuristic technique can resolve all optimization problems satisfactorily. To address various optimization problems, scientists have attempted to develop new metaheuristics or improve existing ones. To overcome the difficulty of constructing DCNN structures without human intervention, we employ powerful natureinspired techniques called Marine Predator Algorithms (MPA) (Mirjalili, 2016) . A novel flexible encoding strategy is proposed in order to circumvent the limitations of fixedlength encodings. As a result, the fundamental objective of the proposed model is to develop and implement a complementary MPA capable of discovering optimal DCNN structures automatically. To summarize, this article makes the following contributions: 1. Develop a novel variable-length MPA method that effectively encodes a DCNN architecture using a unique candid solution encoding technique. 2. Develop a method for learning variable-length DCNN structures independent of the fixed-length encoding constraint imposed by existing MPA; a new layer dubbed Enfeebled will be presented to construct a variablelength predator. 3. Develop a technique for evaluating fitness that uses a portion of the dataset rather than the complete dataset to accelerate the development process. The following is the organization of the paper. Section 2 summarizes the DCNN, MPA algorithms, and the CT Scan datasets. Section 3 discusses the proposed approach for variable-length encoding. Section 4 summarizes and discusses the simulation results. Section 5 is the concluding section. This section describes the background knowledge of the LeNet-5 DCNN, MPA, and CT Scan datasets. DCNN is a straightforward yet successful model among the emerging deep learning-based techniques (Le Cun, 2015) . We chose this network since its structure is straightforward and needs minimal parameters. As illustrated in Fig. 1 , the DCNN is composed of three major components: convolution, subsampling, and fully connected layers. Additionally, feature maps, i.e.,FM k ij and sub-sampling operation, are respectively formulated by Eqs. (1) and (2). where b and are respectively bias values and learning constants, and also n×n i are inputs (Liu et al., 2022a, b) . The last fully-connected layer then performs classification. This layer has one neuron since our problem is a binary task (COVID and non-COVID). In line with most metaheuristic algorithms, MPA is categorized as a population-based algorithm. Thus, the initial solution is regularly disseminated across the searching space as initial testing: where X min and X max signify the lower and upper bound of variables, respectively, and rand represents a uniformly distributed random vector that fluctuates from 0 to 1. It is worth stating that the best solution obtained so far in MPA is appointed as a leading predator for constructing a matrix named Elite, as expressed in Eq. (3). It is also essential to indicate that based on the information of the prey position, these matrix arrays will manage searching and finding the Prey. (1) Here, X I signifies the best searching agent vector reproduced n times for constructing the Elite matrix. In which n represents the search agents count, d represents dimension numbers. The Elite is updated if the best searching agent is replaced with the better agent. Prey is another matrix that has a similar dimension as Elite. Prey is a matrix in which the searching agents update their position. This means that the initial Prey in which the best searching agent creates the Elite is produced by the initialization. The Prey is expressed in Eq. (5). X i,j In Eq. (5), signifies the jth dimension of ith searching agents. It is worth noting that the entire optimization procedure gets primarily and precisely associated with Elite and Prey matrices. The MPA's optimization procedure consists of three stages taking into consideration the various ratio of velocity, and together with simulating the complete life of both predator and Prey, these stages are (I) High-Velocity Ratio (HVR) or once a predator gets moving slower than Prey, (II) Unit-Velocity Ratio (UVR) or once both Prey and predator would have practically the identical speed, and (III) Low-Velocity Ratio (LVR) once the Prey gets moving slower than predator. The above three stages could further be described as follows: Fig. 1 The DCNN architecture • Stage 1: HVR or once the Prey gets moving slower than the predator. This setup takes place in the early iterations of optimization, where the exploration matters. The mathematical model of this rule is expressed as below: Here R BM represents a normally distributed random vector signifying the BM. The symbol ⊗ indicates entry-wise multiplications. In which the multiplication of R BM by Prey mimics the prey motion. P signifies a constant number and its value set to 0.5, and R is a vector of uniform random numbers ranging from 0 to 1. • Stage 2: In this stage, the Prey and predator move very similarly in pace. Here both exploitation and exploration are very crucial. Accordingly, 50% of the searching agents are allocated for exploration, and the remaining 50% are given for exploitation. The predator in this stage is accountable for exploration and the Prey for exploitation. Depending on the rule, in the UVR (v ≈1), the safest policy for a predator is BM if the Prey goes in LF. Accordingly, this research deems predator moves in BM while Prey moves in LF, as Eq. 7. Here R LF represents a random number vector depending on the LF. R LF and Prey multiplication mimic the prey movement in LF, whereas adding the prey position to the step size imitates the prey movement. The MPA supposes the following expressions for the other half of the populations: Here CF represents a parameter that can be adaptive to manage the predator movements step size. R BM and Elite multiplication mimic predator movement in BF, whereas prey modifies its position depending on the predator's movement in BM. • Stage 3: This is where the Prey is moving slower than the predator, which can be introduced as follows: For: the first 1 2 × searching agents One more thing that triggers a change of behavior in marine predators is environmental or ecological problems, for instance, the Eddy formation or Fish Aggregating Devices (FADs) effects (Filmalter et al., 2011) , which is deemed as local optima avoidance operator. Stagnation in local optima can be avoided by considering these longer jumps during the simulation. Therefore, the FADs impact is precisely given as: Here, the value of FADs = 0.2, which is considered the FADs probability impact on the process of optimization. U signifies the binary vector in conjunction with arrays as well as 0 and 1. This is created by producing a random vector ranging from 0 to 1 and modifying its array to 1 if it is greater than 0.2 and 0 if it is smaller than 0.2. r represents the uniform random number, which ranges from 0 to 1. Both X min and X max are considered vectors containing the lower and upper bound of the dimensions. r 1 and r 2 signify random indexes of the Prey matrix. Depending on the emphasized regions, marine predators hold an excellent memory in recapping where they have had successful hunting. The ability is mimicked through memory preservation within the MPA. The Elite matrix will be updated by modifying the Prey and executing FADs impact. The suitability of every searching agent of the existing iteration is assessed to its corresponding in an earlier iteration, and the existing one supersedes the searching agent if it is a better one. COVID-CT and SARS-CoV-2 [20] are two CT Scan datasets used in this study, whose details are discussed in the following subsections. SARS-CoV-2 was established at hospitals in Sao Paulo. This dataset contains 2482 CT scans of 120 patients, 1252 of whom are SARS-CoV-2 patients, and 1230 of whom are not infected but have a variety of respiratory ailments. Typical, COVID-19 positive and normal samples are depicted in Fig. 2 . The first raw contains some COVID-19-infected instances, whereas the second raw contains regular (noninfected) instances. It should be noted that the SARS-CoV-2 has no extraordinary uniformity in terms of contrast and image sizes, as shown in Fig. 3 . The COVID-CT dataset 1 contains CT scans of people infected with COVID-19. This collection includes 349 photos of 216 patients. Two more datasets, LUNA and MedPix, were employed to provide images of non-COVID and healthy individuals. This dataset segment contains 463 photographs of 55 healthy individuals. Similar to the first dataset, the second contains no specified image contrast and size adjustment. As a result, as described in Perumal and Velmurugan (2018), we begin by enhancing the quality and size of the low-contrast images. Fig. Figure 4 illustrates both the unaltered and improved versions. The datasets utilized in this study are listed in Table 1 . In order to evolve a DCNN, this section describes the IPbased MPA (IPMPA) technique in detail. The IPMPA algorithm is structured in Algorithm 1. The population will be initialized in three steps: encoding candid solutions, updating positions, and checking for termination. The IP address structure of the network affects the IPMPA encoding strategy. Three layers comprise the DCNN structure, each with its own parameters. The following table summarizes the various layers. We can establish a network IP address with a fixed length and split it to allow different DCNN layers. The IPMPA uses the IP address as the basis for its encoding technique. Three different types of layers comprise the DCNN architecture: convolutional, pooling, and fully connected. As illustrated in Table 2 , each layer type has its own unique set of values and ranges for parameter settings. A length must first be calculated for an IP-based encoding system. Three key parameters for convolutional layers, namely the number of feature maps, the filter size, and the stride size, are listed in the Parameter column of Table 2 . Second, parameter ranges are adjusted to [1, 8] , [1, 128] , and [1,4], respectively, based on the sizes of the datasets. The decimal values 3, 5, and 3 can be converted to digitized strings 011, 000 0101, and 11, respectively, where the digitized string is finished with zeros until it reaches the appropriate bit length, as illustrated in the example value column of Table 2 . Eventually, the summary row of the convolutional layer in Table 2 presents the total bit count of 12 and the sample digitized string 011 000 0101 01 obtained by connecting the strings of the three previously discussed parameters. The total number of bits and the sample binary string can be obtained using the same technique as with convolutional layers from pooling and fully-connected layers, as shown in Table 2 . Because a layer's maximum bit count is twelve and an IP address is eight bits (one byte), a twelve-bit IP address will require 2 bytes. Table 3 , all DCNN layers must have CIDR-compliant subnets. Since DCNN layers are classified into three types, each type requires three subnets to be completely supported. The length of the IP address 0.0 plus the bit numbers of the convolution layer constitute the subnet mask (12 bits). 0.0/4 (0.0-15.255) is the subnet mask. The initial IP address for the pooling layer is 16.0, calculated by adding one to the IP address for the final convolution layer. Like the convolution layer, the subnet mask has a length of five, resulting in 16.0/5. As a result, subnets 16.0-23.255 correspond to the pooling layer. A similar issue arises with the fully-connected layer subnet 24.0/5, which spans the range 24.0-31.255. The subnets' arrangement is depicted in Table 3 . To adjust DCNN topologies with varying lengths, several layers in the IP-coded candid solution vector are disabled during initialization. To compensate, the Enfeebled layer and subnet are introduced. When all three DCNN layers are used, the Enfeebled subnet has an IP address range of 32.0/5, as demonstrated in Table 3 . A binary string representing a particular layer is padded with zeros until it reaches the length of 2 bytes, at which point the subnet mask is implemented, and the string is converted to an IP address by splitting it by full stops as shown in Table 4 . As a result, the 2-byte digitized strings 0000 0011 and 100 001 can be found in Table 2 . In order to obtain the IP address 6.33, the first byte was converted to 6 and the second to 33. Before computing search agents, each layer must be converted to a two-byte IP address. First, some properties such as maximum length and maximum completely linked length must be stated. These are detailed in Table 5 . The array that stores position data will be twice the size of a single byte, with each byte representing one dimension of the candid solution. Consider the following candid solution vector illustration to better understand DCNN encoding and its variablelength architecture. In the DCNN architecture depicted in Fig. 5 , the Pooling (P), Convolution (C), Fully Connected (F), and Enfeebled (E) layers can all be expressed by IP addresses. The size of the relevant candid solution vector is depicted in Fig. 6 . Following many MPA modifications, the 9th and 10th dimensions of the candid candidate solution may change to 18 and 139, respectively, converting the 3rd IP address indicating an Enfeebled layer to a Pooling layer and so changing the candid solution's DCNN architecture to five layers. In summary, as illustrated in this example, IPMPAencoded searching agents can represent DCNN topologies with various lengths, namely 3, 4, and 5. Population size is determined after initialization and created until the population size is met. Each empty vector of predators contains a network interface, including the individual's IP address and subnet information. Each subsequent layer may contain a convolution, a pooling, or an enfeebled layer. Any four layer types may be employed before the final fully connected layer. The last layer is formed of a layer that is fully connected. Each layer's subnets will be issued a random IP address. Prior to fitness evaluation, a weight initialization method must be chosen, and the Xavier weight initialization approach (Postel, 1980) is chosen since it has been demonstrated to be effective and is utilized in a large number of deep learning systems. For the first half of the training dataset, individuals will be trained using the settings for their previously decoded DCNN architecture (as described in Algorithm 2). After batch evaluating the partially trained DCNN on the second segment of the training dataset, a sequence of accuracy rates is created. Following that, we calculate the candid solution's fitness by averaging the accuracy rate of each participant. The next step is to alter the position of the candid solution by applying the coefficient values for each byte included inside it. Each interface in the candid solution vector should have a unique IP address given to it to ensure that they all have access to the same subnet. The restrictions of each interface, such as the second interface's ability to act as a convolution layer, pooling layer, or enfeebled layer, may change depending on its position in the overall solution vector. Generally, we incorporate IPMPA to increase the DCNN's diagnostic accuracy. For all tests, the number of predators (population size) is limited to 50, and the maximum number of iterations is limited to 200. DCNN is designed to learn at a rate of 11 at a batch size of 0.0002. Additionally, the epochs for each assessment are chosen between 1 and 10, and it is underlined that pictures must be downsampled to 31 × 31 before being fed to DCNNs. All experiments were carried out in MATLAB R2019b on a computer provided with an Intel Core i7-7700HQ processor running at a maximum frequency of 3.8 GHz, Windows 10, and 16 GB RAM. This paper compares IPMPA's performance to that of IPPSO (Wang et al., 2018) , variable-length genetic algorithm (VLGA) (Qiongbing and Lixin, 2016) , variablelength NSGA-II (VLNSGA-II) (Pal et al., 2021) , variablelength brain storm optimization algorithm (VLBSO) (Cheng et al., 2021) , IP-Modified PSO (IPMPSO) (Abbas, 2018) , variable-length biogeography-based optimizer (VLBBO) , and variable-length ant colony optimization (VLACO) (Liao et al., 2013) on the two datasets used in the study. The MPA and other benchmark models' parameters are summarized in Table 6 . This research employs several performance indicators, including sensitivity, accuracy, specificity, F1-score, and precision. Equations 11-15represent the metrics' equations. T N d e n o t e s t h e n u m b e r o f t r u e n e g ative cases, TP denotes the number of true positive cases, FP denotes the number of false-positive cases, and FN denotes the number of false-negative cases. As previously stated, while various resilient deep convolutional neural network designs have been developed recently, we choose the fundamental DCNN architecture due to the computational expense of this task. The first experiment assigns a probability rank to each image in the two datasets, SARS-CoV-2 CT-Scan and COVID-CT, reflecting the likelihood that each image is associated with COVID19 positive cases. This likelihood is then compared to a preset cut-off number in order to determine whether the inputs are infected. Notably, the planned model should obtain a probability of one for infected samples and zero for uninfected samples. Figures 7 and 8 illustrate the EPG distributions for SARS-CoV-2 and COVID-CT. Infected photos, by definition, have a higher probability than uninfected images. The confusion matrices for the SARS-CoV-2 CT-Scan and COVID-CT datasets are depicted in Figs. 9 and 10. The specificity, sensitivity, precision, accuracy, and F1-Score of the DCNN-IPMPA and other benchmarks used in the second experiment, namely IPPSO, VLGA, VLNSGA-II, VLBSO, IPMPSO, VLBBO, and VLACO, are summarized in Tables 7 and 8. The results indicate that DCNN-IPMPA has the highest accuracy score, 97.21 percent for the SARS-CoV-2 dataset and 97.94 percent for the COVID-CT dataset. Additionally, the second-best result obtained utilizing DCNN-VLNSGA-II is 97.31 and 97.00 percent for SARS-CoV-2 CT-Scan and COVID-CT, respectively. Notably, the optimal outcome is highlighted in bold letters. While a variety of metrics can be used to compare pattern recognizers, only one metric can provide a comprehensive view of the performance of benchmark algorithms across a range of thresholds, i.e., precision-recall curve. This diagram illustrates the relationship between recall and precision rate. The precision-recall curves for DCNN-IPMPA and eight benchmarks on the SARS-CoV-2 and COVID-CT datasets are shown in Figs. 11 and 12. The Receiver Operating When comparing metaheuristic algorithms, convergence speed is a critical metric to consider. In addition to the plots discussed above, Fig. 13 depicts the convergence curves of comparison benchmarks to enable comparisons. The above figures show that the DCNN-IPMPA model surpasses all other comparable models. Additionally, it can As previously stated, accuracy rate vs. processing time is a trade-off. In the preceding section, the accuracy of the proposed model was compared to that of various benchmarks. This section compares the processing time of the developed framework to ensure an accurate comparison. On an Intel Core i7-7700HQ CPU and an NVidia Tesla K20 GPU, we built the DCNN-IPMPA and comparison networks. The time values for the 2633 training images and 659 test images are shown in Table 9 . Additionally, the best outcome is highlighted in bold type. Comparing the test and training times of benchmark networks implemented on GPU and CPU is shown in Table 9 . The results of Table 9 reveal that the DCNN-IPMPA outperforms other IP-based benchmarks by 61 s and 215 ms on the CPU and GPU, respectively. By considering the entire amount of processed images, i.e., 3392, it is concluded that DCNN-IPMPA requires less than 0.001 s for each image for testing and training. It must be noted that the DCNN processing time is the best among all benchmarks as expected; however, DCNN-IPMPA represents a competitive result compared to the standard DCNN. This subsection assesses the sensitivity of the DCNN-control IPMPA's parameters. The parameter FADs is deemed a local optima avoidance operator for MPA. The second and third parameters (N Layer and N batch , respectively) address the network structure. The analytical results demonstrate the robustness and sensitivity of the parameters to input. Four different parameter values were used in the corresponding tests (Wu et al., 2021a, b) . Numerous parameter combinations were established using an orthogonal array, as illustrated in Table 10 . The model is trained for all potential parameter combinations. Additionally, Table 11 summarizes the mean square error (MSE) values obtained for each test. Figure 14 shows the trend in the parameter values concerning the data in Table 11 . The result indicates that N Layer = 5, FADs = 0.2, and N batch = 10 produce the best results. This paper evaluates the MPA's effectiveness in determining the best structure of DCNNs without requiring manual effort to achieve both speed and accuracy. Three enhancements based on standard MPA were developed to accomplish the purpose. First, a novel encoding technique The convergence curves for utilized benchmark algorithms based on the IP address was proposed; then, an Enfeebled layer was presented to cover the dimensions of specified candid solution vectors. Finally, the learning process partitioned enormous datasets. The DCNN-IPMPA was applied to the SARS-CoV-2 and COVID-CT datasets to conduct a detailed comparison. The performance of the DCNN-IPMPA was compared to that of standard DCNNs, and DCNNs evolved using IPPSO, VLGA, VLNSGA-II, VLBSO, IP-MPSO, VLBBO, and VLACO using five well-known metrics: sensitivity, accuracy, specificity, F1-Score, and precision, as well as ROC and precisionrecall curves. The study established that the presented system outperforms competing models, achieving a final accuracy of 97.21 percent on the SARS-CoV-2 dataset and 97.94 percent on the COVID-CT dataset. Timing analysis indicated that the DCNN processing time is the best among all benchmarks as expected; however, DCNN-IPMPA represents a competitive result compared to the standard DCNN. Numerous directions for future research study exist, including the application of DCNN-IPMPA to various image processing tasks. Researchers can explore the IPMPA in order to address multi-objective optimization problems. Additionally, a unique fitness function can be built to more accurately model the issue. As additional study directions, the use of oppositional-based learning, chaotic maps, orthogonal learning, and Gaussian walks can be investigated to optimize DCNN performance. Also, other DCNN structures or vision transformers can be used for future research studies. Availability of data and material The resource images can be downloaded using the following links: https:// github. com/ UCSD-AI4H/ COVID-CT; http:// medic alseg menta tion. com/ covid 19/ Internet protocol (IP) steganography using modified particle swarm optimization (MPSO) algorithm COVID-CT-MD, COVID-19 computed tomography scan dataset applicable in machine learning and deep learning Correlation of chest CT and RT-PCR testing in coronavirus disease 2019 (COVID-19) in China: a report of 1014 cases SARS-CoV-2 CT-scan dataset: a large dataset of real patients CT scans for SARS-CoV-2 identification Chest CT findings in coronavirus disease-19 (COVID-19): relationship to duration of infection Interhuman transmissibility of Middle East respiratory syndrome coronavirus: estimation of pandemic risk Multiobjective evolution of the explainable fuzzy rough neural network with gene expression programming Adaptive CCR-ELM with variable-length brain storm optimization algorithm for class-imbalance learning CT imaging and differential diagnosis of COVID-19 First descriptions of the behavior of silky sharks, Carcharhinus falciformis, around drifting fish aggregating devices in the Indian Ocean Disease staging and prognosis in smokers using deep learning in chest computed tomography Rapid ai development cycle for the coronavirus (covid-19) pandemic: Initial results for automated detection and patient monitoring using deep learning ct image analysis Radnet: radiologist level accuracy using deep learning for hemorrhage detection in ct scans Sample-efficient deep learning for covid-19 diagnosis based on ct scans Real-time COVID-19 diagnosis from X-Ray images using deep CNN and extreme learning machines stabilized by chimp optimization algorithm Real-time COVID-19 diagnosis from X-Ray images using deep CNN and extreme learning machines stabilized by chimp optimization algorithm Study of the relationship between pilot whale (Globicephala melas) behaviour and the ambiguity function of its sounds Passive sonar target classification using multi-layer perceptron trained by salp swarm algorithm Chimp optimization algorithm Classification of underwater acoustical dataset using neural network trained by Chimp Optimization Algorithm Evolving deep learning convolutional neural networks for early COVID-19 detection in chest X-ray images 14 The results of the control parameters LeNet-5, convolutional neural networks 20:14 Ant colony optimization for mixed-variable optimization problems A survey of deep neural network architectures and their applications Composition and divergence of coronavirus spike proteins and host ACE2 receptors predict potential intermediate hosts of SARS-CoV-2 Federated neural architecture search for medical data security Improved feature point pair purification algorithm based on SIFT during endoscope image stitching Position-transitional particle swarm optimization-incorporated latent factor analysis SCA: a sine cosine algorithm for solving optimization problems Training a feed-forward neural network using particle swarm optimizer with autonomous groups for sonar target classification Neural network trained by biogeography-based optimizer with chaos for sonar data set classification Rajendra Acharya U (2020) Automated detection of COVID-19 cases using deep neural networks with X-ray images Portfolio formation and optimization with continuous realignment: a suggested method for choosing the best portfolio of stocks using variable length NSGA-II DoD standard internet protocol Underwater targets classification using local wavelet acoustic pattern and multi-layer perceptron neural network optimized by modified whale optimization algorithm A new crossover mechanism for genetic algorithms with variable-length chromosomes for path optimization problems Face recognition by using neural network Design of a fuzzy model of control parameters of chimp algorithm optimization for automatic sonar targets recognition Diagnosis of COVID-19 using CT scan images and deep learning techniques Using deep learning for classification of lung nodules on computed tomography images Evolving neural networks through augmenting topologies A genetic programming approach to designing convolutional neural network architectures Evolving deep convolutional neural networks by variable-length particle swarm optimization for image classification Training RBF NN using sine-cosine algorithm for sonar target classification No-Free-Lunch Theorem Evolving deep convolutional neutral network by hybrid sinecosine and extreme learning machine for real-time COVID19 diagnosis from X-ray images Acoustic detection and recognition of dolphins using swarm intelligence neural networks A multilayered-and-randomized latent factor model for high-dimensional and sparse matrices Endoscope image mosaic based on pyramid ORB Deep learning-based detection for COVID-19 from chest CT using weak label The authors declare that there is no conflict of interest.