Application of a novel deep learning technique using CT images for COVID-19 diagnosis on embedded systems

Problem A novel coronavirus (COVID-19) has created a worldwide pneumonia epidemic, and it's important to make a computer-aided way for doctors to use computed tomography (CT) images to find people with COVID-19 as soon as possible. Aim: A fully automated, novel deep-learning method for diagnosis and prognostic analysis of COVID-19 on the embedded system is presented. Methods In this study, CT scans are utilized to identify individuals with COVID-19, pneumonia, or normal class. To achieve classification two pre-trained CNN models, namely ResNet50 and MobileNetv2, which are commonly used for image classification tasks. Additionally, a novel CNN architecture called CovidxNet-CT is introduced specifically designed for COVID-19 diagnosis using three classes of CT scans. To evaluate the effectiveness of the proposed method, k-fold cross-validation is employed, which is a common approach to estimate the performance of deep learning. The study is also evaluated the proposed method on two embedded system platforms, Jetson Nano and Tx2, to demonstrate its feasibility for deployment in resource-constrained environments. Results With an average accuracy of %98.83 and an AUC of 0.988, the system is trained and verified using a 4 fold cross-validation approach. Conclusion The optimistic outcomes from the investigation propose that CovidxNet-CT has the capacity to support radiologists and contribute towards the efforts to combat COVID-19. This study proposes a fully automated, deep-learning-based method for COVID-19 diagnosis and prognostic analysis that is specifically designed for use on embedded systems.


Introduction
On December 31, 2019, a virus called SARS-CoV2 was first identified in the city of Wuhan, located in China. This virus is responsible for causing illnesses related to the coronavirus, also known as COVID-19. Since its initial discovery, this virus has rapidly spread across the globe, affecting populations in var-ious countries and regions. [1,2]. The World Health Organization (WHO) classified the COVID-19 outbreak as a pandemic following this event. They stated that techniques, processes, and resources would be required to promptly identify people most at risk of impairment and death. The impacts of COVID-19 vary among individuals. However, it is worth noting that several infected patients endure mild to moderate illness and recuperate without hospitalization. [3,4]. Fever, dry cough, and exhaustion are the most prevalent symptoms, and they usually start mildly in all patients. It is important to note that a small percentage of individuals may experience severe symptoms such as chest pain or pressure, difficulty speaking or moving, and shortness of breath. These symptoms should not be ignored and require immediate medical attention [5,6].
The severity of this disease can vary significantly, ranging from a mild, self-restricting respiratory ailment to a serious and advancing pneumonia, resulting in multi-organ malfunction and eventual fatality [6][7][8][9]. As the pandemic continues to advance, the implications of this viral disease are causing a great deal of anxiety in society. With an increasing count of confirmed cases, patients are experiencing severe respiratory collapse and cardiovascular complications, leading to heightened concern [10]. Although there are no particular therapies or vaccinations for COVID-19, there are several current clinical trials examining possible remedies. Regardless of immunization or treatment deficits, people may prevent illness by cleaning their hands, remaining at home, coughing or sneezing into their hands, and not smoking. These precautions are not therapeutic, but they can help individuals avoid disease and decrease the spread of COVID- 19.
The increasing number of confirmed COVID-19 cases and the rising incidence of severe respiratory and cardiovascular problems have sparked concern over the virus's impact. As the pandemic continues to spread, it is becoming increasingly clear that the virus can have significant and potentially dangerous consequences [10]. The discovery of effective ways for resolving COVID-19-related difficulties has gotten a lot of attention. In the battle against the virus, researchers and decision-makers are facing a significant challenge -dealing with the vast amount of data, commonly known as big data. This surge in information is putting a strain on their resources, and they must find ways to manage and utilize this data effectively. This exemplifies the extent to which Artificial Intelligence (AI) can be involved in enhancing and advancing global healthcare systems [11]. Computer vision methods, which were originally used for categorization of general images, have been applied to medical images, including computed tomography (CT) scans, thanks to the rapid growth of artificial intelligence [12]. The convolutional neural network (CNN) is a frequently used feedforward neural network, and various models have shown considerable promise in capturing feature representations among the available methodologies [13,14]. Previous research has shown that individuals with COVID-19 display abnormalities in the form of ground glass opacities when examined using chest CT scans. Researchers suggest that a technique that relies on chest CT scans could be an effective method for identifying and quantifying cases of COVID-19. This approach has the potential to be a valuable tool in the fight against the virus [15]. Using X-ray imaging, a number of researchers have proven several methods for identifying COVID-19. Computer vision [16], machine learning [17][18][19] and deep learning [20][21][22] have recently been utilized to autonomously identify numerous disorders in the human body [23], allowing for intelligent healthcare.
Consequently, the employment of X-ray imaging, radiography, and computed tomography (CT) techniques has become crucial in the prompt detection and diagnosis of COVID-19 [24][25][26][27]. Radiological results have demonstrated that CT has a high diagnostic and prognostic value for COVID-19 in recent investigations. In the diagnosis of COVID-19, for example, CT had a substantially better sensitivity than RT-PCR [28,29]. On CT images of COVID-19 patients, bilateral lung lesions comprised of ground glass opacities were common [29,30]. Even asymptomatic individuals had abnormalities and alterations on repeated CT scans [31]. CT is a popular diagnostic technique that is simple to get and does not come at a high cost. Using CT imaging to provide a precise diagnostic tool helps speed up the diagnosis process and is a good complement to RT-PCR. The use of CT imaging to predict an individual's customized prognosis could prove beneficial in identifying those at high risk of severe illness, who require urgent medical attention. By doing so, medical professionals can ensure timely and appropriate treatment for those who need it most. Therefore, this approach may be a crucial step in managing and combating the impact of COVID-19 [32][33][34]. Mahmoudi et. al. present a deep learning-based diagnosis system for the detection and quantification of COVID-19 infection and pneumonia screening using CT imaging [35]. The proposed system in the study is a U-network architecture for image segmentation and a three-layer CNN architecture for classification. The lung and infection segmentation models achieved a high membrane score of 0.98 and 0.91, respectively, with an accuracy of 0.98 for the overall classification. The proposed system shows potential for automated detection of lung infections from CT scans, even with limited training datasets. Ibrahim et. al. proposed a hybrid deep learning approach to accurately identify COVID-19 patients based on their lung CT images [36]. The proposed system consists of a lung segmentation technique, a post-segmentation method, and three pre-trained deep learning models that were combined to increase the overall prediction capabilities. The proposed model achieved a 95% accuracy rate on a publicly available COVID-19 CT dataset, outperforming several state-of-the-art studies. The developed model could serve as an additional diagnostic tool for leading clinical professionals. Abdulkareem et. al. developed a COVID-19 diagnostic system using a convolutional neural network, stacked autoencoder, and deep neural network [37]. The proposed system outperformed current state-of-the-art models in detecting the COVID-19 virus using CT images, achieving an accuracy rate of 88.30%. A large-scale and challenging CT image dataset was used to train the deep learning model, and its final performance was reported. A new selection scheme is proposed in this study using a crow swarm optimization algorithm to select the best deep learning model for COVID-19 diagnosis by Mohammed et. al. [38]. Two datasets are used to evaluate the proposed method, and ResNet50 and VGG16 are found to be the optimal deep learning models for the respective datasets. Among the pretrained CNN and deep models using the first dataset, ResNet50 has an accuracy of 91.46% and a F1-score of 90.49%. This methodology can assist healthcare managers in selecting and evaluating the best COVID-19 diagnostic models based on deep learning. Song et al conducted a study using a deep learning model to detect patients infected with the COVID-19 virus [39]. With this model, which they introduced as DRE-Net, they tried to distinguish between COVID-19 patients and bacterial pneumonia patients. In the study, they reported that they were able to accurately distinguish patients infected with the COVID-19 virus from other patients. As a result of the study, they stated that they obtained a precision rate of 0.93 and an accuracy rate of 0.93 for three types of CT. Khobahi et al. developed a semi-supervised deep learning system based on automatic encoders called CoroNet to detect patients infected with COVID-19 [40]. The proposed system combined three open access datasets for experiments. In this scheme, 18,529 images in different categories were used. Overall, accuracy and F1-score were obtained from the experiment 93.50% and 93.51%, resprectively. The study submitted by Allioui et. al describes a new method for mask extraction in medical image segmentation, that is based on multi-agent deep reinforcement learning [41]. This method aims to improve the efficiency of automatic image segmentation and minimize the need for manual mask extraction. The proposed method is tested on COVID-19 CT images and achieved high precision, sensitivity, and specificity, making it a promising tool for accurate clinical diagnosis of COVID-19. The results demonstrate the effectiveness of using deep reinforcement learning for CT mask extraction.
As the applicability of deep learning continues to expand, researchers have focused on developing more efficient ways to implement these techniques on embedded platforms or single board computers. Numerous studies have investigated this area, and their findings suggest that while the Raspberry Pi is an energy-saving option, NVIDIA Jetson platforms such as the Jetson Nano, Jetson TX1, and Jetson TX2 offer better performance due to their faster GPUs. This information could be valuable for those seeking to optimize their deep learning applications on these types of devices [42]. When comparing the performance of Jetson with that of the Raspberry Pi, Jetson comes out on top [43]. Because it offers developer kits with a range of functionalities, the NVIDIA Jetson series is a pioneer in single-board computing for deep learning applications. Energy economy, lightweight and portable design, ease of use, and excellent performance per watt are just a few of the Jetson series' advantages. In the machine learning extraction phase, NVIDIA's Jetson series is the most often utilized accelerator [42]. A block diagram of the study is given in Fig. 1.
In this study, a new fully automated deep learning method for diagnosis and prognostic analysis of COVID-19 on embedded system is presented. The study performs the classification process in three classes, and it is portable with the integration of the system into embedded systems. In this way, it creates a useful decision support mechanism with adequate infrastructure in areas where health services are lacking. In addition, the main contributions of the study are given below.
The use of CT scans to diagnose and classify COVID-19, pneumonia, and normal cases is crucial in the fight against the ongoing pandemic. The lack of three-class classification studies, especially in the diagnosis of COVID-19, increases the importance of the study. This research is significant because it presents a comparative analysis of a newly proposed CNN model and two pre-trained architectures from the literature, providing insights into the effectiveness of different models in diagnosing and categorizing COVID-19. The proposed CNN model is smaller in size and therefore takes up less space. The number of parameters is less and requires less computational power. In addition, a higher accuracy value is reached.
To ensure the reliability of the results, the study utilizes the k-fold cross-validation technique, a widely used method for evaluating deep learning models. The portability of the proposed model is also demonstrated by testing it on different embedded system boards, and the training and testing times of these systems are presented, providing valuable information for the development of practical applications. Finally, to put the research into context, a comparison table is created with studies from the literature, highlighting the novelty and contributions of this work.
The study's structure is given below. In Section 2, a basic overview of the dataset and deep learning is provided, as well as a detailed explanation of the proposed network model. Furthermore, details on the parameters and embedded systems for the proposed deep learning application models are provided. Section 3 presents the results of evaluating deep learning classifiers on embedded systems, including key metrics such as accuracy, recall, precision, and F1-scores. Additionally, the findings of cross-validation methods are also discussed. This section provides valuable insights into the performance of deep learning algorithms when applied to embedded systems. The discussion section is given in Section 4, and the conclusion section is submitted in Section 5.

Deep learning
The convolutional neural network's deep learning structure is a more advanced version of the convolutional neural network. The attributes of the input data are automatically retrieved and transferred to the following layers using these qualities throughout the convolution process. Convolution is a tech- nique for extracting different features from an image. The folds on the edges appear to be the most prominent points. A different procedure is carried out in each layer of the deep learning convolutional neural network, which has a multilayer structure, and data is passed to the next layer. Each layer has a specific purpose. The following are the layers of deep learning architectures and the operations performed by these layers.
The CNN architecture is the foundation of the deep learning idea. Fig. 2 depicts the CNN architecture. According to this architecture, the first few steps are made up of convolution and pooling layers. The fully connected layer is followed by the classification layer in the final stage. In summary, CNNs are made up of a series of trainable elements arranged in a row. It is followed by a classifier for instructional purposes. The training procedure is carried out by layer-by-layer processes when the input data is received in CNN [44].

Proposed model
This section of the analysis focuses on the CovidxNet-CT model, a convolution-based algorithm developed to detect COVID-19 from CT scan images. The performance and effec-tiveness of this model are thoroughly analyzed and discussed, providing insights into its potential as a tool for identifying cases of COVID-19. The CovidxNet-CT model is designed using the sequential model of the Keras library. In order to reduce overfitting situations that may occur, dropout layers with a density of 0.3 are added to the model. ReLu activation function, which is frequently used in image processing, is used as the activation function. The layers of the model are given in Fig. 3.
The network architecture consists of the convolution layer, maximum pooling, dropout, and a classification layer. A set of convolution and max pooling layers acts as feature extractors. The network architecture consists of two 3x3 convolution layers with 32 and 64 units, respectively, and two maximum pooling layers of 2x2 sizes. There is also a 30% dropout layer. CovidxNet-CT network model structure parameters are shown in Table 1.

Dataset
Diagnosing a disease can often feel like a glimmer of hope in an otherwise difficult situation. In the case of the COVID-19 pandemic, early identification and detection are of paramount  importance. Therefore, the focus should be on gathering the necessary data to effectively train models for disease detection. This data collection should be given top priority, as it forms the foundation for successful diagnosis and treatment. This information will aid in the diagnosis of COVID-19 cases using machine learning or deep learning methods. For this experiment, a CT scan dataset is used to detect COVID-19 disease [45]. The COVID-CT-MD dataset is accessible through ''https://github.com/ShahinSHH/COVID-CT-MD" (access time, 14.01.2023).
A dataset comprising COVID-19 CT scans, designated as COVID-CT-MD, has been created to facilitate investigations related to the classification of COVID-19 through the use of machine learning and deep learning methods. There are 169 confirmed positive COVID-19 cases in the COVID-CT-MD dataset, 76 normal cases, and 60 community acquired pneumonia (CAP) cases. The data in Dicom format is converted to PNG format with a size of 224 Â 224. An example of each from the COVID-19, Non-COVID-19 pneumonia, and normal classes is given in Fig. 4.
There are a total of 305 cases in the dataset, including 169 COVID-19, 76 normal and 60 pneumonia cases. The number of images per case varies and includes at least 108 images. In this study, a total of 60 cases belonging to each class are used to create a balanced dataset. Undersampling is done for the images of the cases, and a dataset is created by obtaining random images from each case with the help of the algorithm used. To train and test the model, the dataset is split into two groups, with 80% of the data allocated for training purposes and the remaining 20% reserved for testing. This approach ensures that the model is trained on a diverse range of data and can accurately identify and classify COVID-19 cases. The testing data serves as a means of evaluating the model's effectiveness and ensuring that it performs well on previously unseen data. There are 9005 images in total, 7204 for training and 1801 for testing. The number of cases and images in the dataset is given in Table 2.

Pre-Processing
Preprocessing is a common procedure in computer vision applications that involves techniques for improving the quality of images. These techniques can help to reduce unwanted noise, highlight key features of an image, and even facilitate the deep learning training process. In this study, a simple method of pixel density normalization within the range of [0,1] is employed. This preprocessing step is essential for ensuring model convergence during the training phase and improving the accuracy of the final results. The input images for CNN models are frequently resized to keep them compatible with network structures. The suggested model has a low computational cost in terms of latency and memory, allowing higherresolution input images to be used. In order to utilize the CT images for deep learning, they are first compiled into a unified dataset and then resized to a standard dimension of 224 Â 224 pixels. Labels associated with the image data are then used to classify each image as positive for COVID-19, pneumonia or healthy. This process enables the deep learning model to accurately identify and classify COVID-19 cases, which is critical for effective disease detection and diagnosis.

Embedded system
There are several important factors that must be taken into account when implementing artificial intelligence (AI) in embedded systems for practical purposes. These include accuracy, speed of computation, compact model size, and energy efficiency. To meet these requirements, the NVIDIA Jetson Nano GPU is an excellent candidate for testing and evaluation due to its affordability and widespread use in fields such as intelligence, surveillance, and space exploration. Utilizing the potential of the Jetson Nano, a robust yet compact GPU, and researchers can formulate more efficient AI systems that can be implemented across a diverse range of real-world scenarios. The Jetson Nano consists of 128 Maxwell GPU cores, 4 ARM Cortex-A57 CPU cores, 4 GB of volatile DDR4 RAM, a micro-SD card based on flash technology for non-volatile storage, and consumes only 10 W of power [46]. Algorithms, according to artificial intelligence, utilize extremely little electricity. Jetson TX2, the medium-sized board in the Nvidia Jetson ecosystem, is larger than Jetson Nano. Fig. 5 shows the Jetson Nano embedded system that is used.

Experimental results
In addition to the CovidxNet-CT model proposed to diagnose COVID-19 in this section, two pre-trained network architectures named MobileNetv2 and ResNet50, which are previously validated using the ImageNet dataset, are used in the study. The obtained dataset is given to the algorithms after the relevant image processing steps. To train and test the model, the dataset is divided into two groups, 80% of the data is reserved for training purposes and the remaining 20% for testing. The training phase of the classification process carried out within the scope of this study is carried out on the hardware platform given in Table 3 below. Python 3.8.6 programming language, Keras 2.4 and Tensorflow 2.2 versions are used for the classification process. The test phase of the study is carried out using PowerShell on Jetson TX2 and Jetson Nano embedded systems.
This approach allows the model to be trained on a wide variety of data and be able to accurately identify and classify COVID-19 cases. Training parameters are given in Table 4. Total parameters of used architectures are shown in Table 5.
The optimization algorithm of these three models is chosen as Adam and the initial learning rate is 0.0001. The performance metric is accuracy and the loss function is categori-cal_crosentropy. Each model training data is trained in 20 epochs in the form of 100 batch sizes. Then, the trained model is tested with the test data.    Cross-validation is a technique that is utilized to assess the accuracy and reliability of a network model in a statistical resampling approach. This method aims to evaluate the performance of the model objectively and accurately by testing it on different subsets of the data [49]. K-fold cross validation is most commonly used in applied machine learning to forecast a model's performance on data that is not utilized during the model's training [50]. To improve the reliability of the experimental results, a cross-validation method called k-fold cross-validation is used in this study. Specifically, a 4-fold cross-validation technique is employed to train the CNN classifier. For k = 4 fold cross validation, the accuracy of CovidxNet-CT is superior to that of other classification models in terms of accuracy, recall, precision, and F1-score performance measures. In the testing of CovidxNet-CT, Mobile-Netv2 and ResNet50 models, 98.83%, 88.86%, and 90.09% accuracy are obtained, respectively. The classification estimation results for each model are presented in the confusion matrix and classification reports given in Figs. 6-8. Table 6 presents the evaluation metrics for the models trained on each infection type in the COVID-CT-MD dataset using the 4fold cross-validation method. It is observed that the CovidxNet-CT network model achieves both a high positive predictive value and a high F1-score. A 4 fold method can successfully sidestep the overfitting problem, according to crossvalidation research. Compared to MobileNetv2, it is observed that the CovidxNet-CT and ResNet50 models achieve 9% and 1% higher sensitivity, respectively.
According to the other two models used, our CNN model proposed within the scope of the study; It is smaller in size and thus takes up less space. The number of parameters is less. It requires less computational power. Higher accuracy value is reached.
The network model's effectiveness is evaluated using several metrics, such as precision, recall, F1 score, accuracy, and the Area Under the Curve (AUC). The CovidxNet-CT model is assessed with a 99% confidence interval for AUC, and a comparative analysis is conducted to determine the most optimal model. Precision is computed as the ratio of the true positive cases predicted to the total predicted true positive and false positive cases. Recall, on the other hand, is calculated as the ratio of the predicted true positive cases to the sum of predicted true positive cases and false negative cases. Receiver Operating Characteristic (ROC) curves are plotted to assess the model's ability to distinguish between normal and abnormal, COVID-19, and pneumonia in CT images. AUC is used to summarize the diagnostic accuracy of each parameter, with the model's performance improving as the AUC score increases. The AUC score ranges from 0.5 to 1, with a higher score indicating superior performance. The AUC is an important metric that summarizes the diagnostic accuracy of the model, and it is used in the study to assess the CovidxNet-CT model's performance with a 98.83% confidence interval.   Application of a novel deep learning technique using CT images for COVID-19 diagnosis on embedded systems recall. A high AUC of 0.93 is obtained from the ResNet50 model. ROC graphs for four folds of three models are given in Figs. 9-11, respectively. The ROC curves are plotted to visu-alize the model's performance in distinguishing between different classes of CT images, and the AUC score is used to determine the best-performing model. By evaluating the     Application of a novel deep learning technique using CT images for COVID-19 diagnosis on embedded systems model's performance using multiple metrics, the study provides a robust analysis of the CovidxNet-CT model's effectiveness, which can guide its future use in clinical settings. In addition, the accuracy, precision, recall, and F1-score values obtained as a result of 4-fold cross validation for the three models are taken separately for each fold and presented in the graphs in Figs. 12-14, respectively. The relationship between accuracy and loss is typically inversely proportional, as shown in the graphs. Lower loss levels are associated with higher accuracy values. In the network design, the epoch parameter is set so that the loss stays low and does not get worse over time.
The models are trained until they reach stability in order to ensure the highest level of accuracy, as continuing past that point does not produce further gains. As shown in Fig. 15, both models are found to be stable when the number of epochs is less than or equal to 20. Table 7 compares the proposed model to various pre-trained models and shows how the proposed model performs better in terms of more accurately detecting COVID-19.
In particular, the studies made using the dataset that we have benefited from in the literature are given in Table 8. When the table is examined, the accuracy result of our proposed model is higher than other studies, showing the success of our study. Additionally, inside the intended system, a mobile interface platform has been established, and real-time applications may be analyzed easily and rapidly using the mobile application. Table 9 shows the performance of the Jetson Nano and Jetson TX2 in real time. After evaluating the test results, it became evident that the proposed new model outperformed the other models used in the study. In particular, the performance of the Jetson TX2 platform is noteworthy, as it is able to handle the computational demands of the models while maintaining high accuracy. The Jetson TX2 platform's compact size and low power consumption make it a promising tool for implementing AI in real-world scenarios, where efficient and reliable performance is crucial. The results of this study suggest that the proposed model, in combination with the Jetson TX2 platform, can play a significant role in the early

Discussion
We demonstrate that our pre-trained model and new CNN model on the proposed approach have a high degree of generalizability with respect to CT images by applying deep learning to the COVID-19 test using CT images. The proposed model detects COVID-19 cases with a 99 percent accuracy rate. Our proposed method outperforms the pre-trained architecture search model across all of our defined performance metrics. The proposed technique is also evaluated on two embedded system platforms, Jetson Nano and Tx2, to demonstrate its applicability in environments with limited resources. Comparing the test and prediction times of the models on embedded system platforms, it is seen that the proposed new model decides faster than other models. This indicates that our method is capable of accurately identifying COVID-19 negative   Our study patients as negative in the majority of cases, thereby reducing the danger of misdiagnosing COVID-19 negative cases as positive and reducing the burden on the healthcare system. However, it is essential to note, that our study has limitations and has only been validated using a small number of COVID-19 samples. As this outbreak is ongoing, there are currently no large datasets available to the public. In the future, we plan to evaluate our method with large datasets. Future efforts will be made to derive more useful information from CT images in order to the severity of COVID-19 and fight the pandemic. It will be possible to conduct detailed analyses of models, casting light on the COVID-19 detection mechanism, identifying the fundamental characteristics of CT images, and aiding clinical doctors in scanning. Even though the proposed system performs well on publicly available datasets, the work is still in the theoretical research phase, and the models have not been tested yet in actual clinical practice. The study assesses the proposed method on two embedded system platforms, Jetson Nano and Tx2, to demonstrate its applicability in environments with limited resources. However, the technical limitations of these platforms may hinder the efficacy of the proposed method in actual environments. Consequently, we will be able to enhance the proposed models for future studies. Depending on the employed optimization techniques and hyperparameter values, the success of the CNN model and other architecture models can be improved. Future research utilizing models trained on more atypical and ambiguous chest X-ray images may also be able to resolve the issue of classification performance in atypical and ambiguous appearances. The success of these studies will facilitate the use of systems supported by deep learning for both the early diagnosis of disease and the global fight against potential new pandemics.

Conclusion
This study demonstrates the potential of using a deep learning strategy to assist doctors in identifying COVID-19 patients and automatically detecting possible lung lesions on CT scans. The proposed CNN model (CovidxNet-CT) has %98.83 accu-racy and its applicability in diagnosing COVID-19 has been demonstrated. Furthermore, classification is made using two different architectures (MobileNetv2 and ResNet50) in the literature and the results are presented comparatively. Two other CNN models (MobileNetv2 and ResNet50) are also evaluated, and the proposed model outperformed them in terms of both accuracy and speed. The use of a 4-fold cross-validation approach successfully addressed the overfitting problem and increased the reliability of the results. The high accuracy of the proposed model in detecting and evaluating lung radiological images makes it a valuable tool for the early detection and diagnosis of COVID-19 and can assist healthcare professionals in managing the disease. Additionally, the portability of the proposed system, which is tested on two embedded system platforms, Jetson Nano and Jetson Tx2, makes it advantageous in terms of practicality and ease of use. This work contributes to the limited studies in the area of three-class classification of lung radiological images, and the results obtained indicate the potential of the proposed model to become one of the most accurate diagnostic tools for COVID-19.
Limitations of this study include the relatively small dataset size, which may affect the generalizability of the proposed model to larger populations. Additionally, the study only focused on three-class classification for COVID-19, pneumonia, and normal lung scans, and did not include other potential lung pathologies. Future work can include expanding the dataset to increase the diversity of cases and improve the generalizability of the model. Further research can also investigate the model's performance on other lung pathologies and develop a multi-class classification system. Finally, the proposed system's clinical efficacy can be evaluated through a larger-scale clinical trial to assess its utility in real-world clinical settings. The findings of this study can equip radiologists with the ability to quickly and accurately diagnose and predict COVID-19 cases, thereby reducing the negative effects of the disease in all areas. The system developed in this study has the potential to be a valuable tool for the early detection and diagnosis of COVID-19, and further development of this sys-