Challenges issues and future recommendations of deep learning techniques for SARS-CoV-2 detection utilising X-ray and CT images: a comprehensive review

Relevant review articles

This article explores the domain of deep learning techniques used for analyzing SARS-CoV-2. Within the realm of deep learning-based analysis for SARS-CoV-2, numerous reviews and survey publications are available. Table 1 outlines recent deep learning-based review articles centred around SARS-CoV-2 analysis. A review article spotlighted the latest advancements in AI methodologies for detecting SARS-CoV-2 using chest radiography and computed tomography; however, it falls short in addressing limitations and research gaps (Aslani & Jacob, 2023). Another recent survey showcased the development of deep learning techniques, comprehensively illustrating diverse methodologies and their performance. Yet, it does not delve into research gaps (Subramanian et al., 2022). In the year 2023, an article (Ait Nasser & Akhloufi, 2023) presented a review detailing 22 publicly available datasets employed for automatic chest disease detection using radiological medical images and DL techniques. It elucidated preprocessing and data-augmentation techniques, engaged with research challenges, and explored alternatives. Nonetheless, this work does not explicitly delineate key findings and intricate mechanisms. A review article discussed the gamut of deep learning models for coronavirus (SARS-CoV-2) epidemic analysis with technology and application. Regrettably, it lacks a robust methodological exposition (Awassa et al., 2022). Shorten, Khoshgoftaar & Furht (2021) contributed a novel review study on SARS-CoV-2 analysis employing deep learning. This article explored how artificial neural networks feed data to deep neural networks, construct learning paradigms, tackle tasks, and analyze novel datasets, including the utilization of annotation techniques. However, this review fell short in adequately addressing recent limitations, research gaps, findings, and discoveries. Syeda et al. (2021) presented another review, examining novel machine learning techniques in SARS-CoV-2 investigation, scrutinizing recent literature that employed AI techniques to trace, comprehend, and combat virus inflammation. Despite its contribution, this work lacks comprehensive methodological descriptions and research gap identification. Sharma & Guleria (2024) contributed a review article categorically summarizing findings, yet it did not extensively explore research gaps. A review study furnished a decent examination of recent methodologies for SARS-CoV-2 analysis, albeit methodological explanations were lacking (Subramaniam et al., 2023). Manigandan et al. (2020) compiled an inclusive review encompassing recent trends in SARS-CoV-2 transmission, diagnosis, prevention, and imaging features. Regrettably, this publication did not sufficiently elucidate research gaps and methodological intricacies. Another review expounded on the use of deep learning techniques in analyzing SARS-CoV-2 within clinical practice, specifically employing CT images. However, it only examined the results of 10 convolutional neural networks and outlined 10 approaches without addressing gaps in the literature (Ardakani et al., 2020). A review study centered on SARS-CoV-2 estimation and forecasting provided a succinct overview, yet it did not accentuate the drawbacks of each method and focused primarily on the outcomes of seven models (Salehi, Baglat & Gupta, 2020).

By and large, many of these studies employ traditional methods to dissect contemporary SARS-CoV-2 research. Existing survey articles often provide concise explanations to illustrate methodologies and related works in a generalized manner. A recurring pattern is observed in traditional survey articles for SARS-CoV-2 research, frequently encompassing the description of the compared samples or related studies.

There are a few essential features to look for when judging a successful review article. We have compiled a list of the top fifteen characteristics that guarantee a high-quality review article. If any literature reviews fail to report a few of the review system’s characteristics, then such articles suggest that some aspects of the review system are missing.

Table 1 shows the limitations and findings of the existing review based on our selected fifteen characteristics of a good review article.

Our comprehensive survey article thoroughly investigates the SARS-CoV-2 study’s classification system, prevailing architecture, structures, and dataset utilization. This encompasses a succinct overview of data attributes, challenges, performance metrics, output evaluations, and prominent recent research citations, shedding light on the foundational equations that underlie deep learning algorithms.

In the critical review segment of our report, we expound on analogous research endeavours, providing specific insights into their methodologies, datasets, the quantity of class predictions, performance results, findings, and areas of research gaps. This critical assessment spans the entire spectrum of deep learning-based SARS-CoV-2 analytical techniques, offering a concise summary based on specified attributes.

Our main contribution to this review study can be presented as follows:

• To investigate systematically on state-of-the-art deep learning breakthroughs on imaging models in SARS-CoV-2 medical diagnostics, including categorization, recent applications, methodology, data, and obstacles.

• Discuss the overall findings, performances, uses, and research gaps, together with pertinent recent research references, basic datatype operations, deep learning models with fully illustrated structure and formulas, and the SARS-CoV-2 evaluation algorithm.

• Critical comparison of all modern SARS-CoV-2 analysis methods based on deep neural networks including advantages, weaknesses, limitations, research directions, and concern about some of the most successful methods.

• Experimentally analyze the twelve most sophisticated deep learning techniques for SARS-CoV-2 analysis, with a focus on the most useful technique.

• To discuss the challenges, privacy, benefits and drawbacks of deep learning models and recommend the best deep learning tools for successful SARS-CoV-2 detection.

The arrangement of the subsequent sections in this study is detailed in Fig. 1. The initial segment furnishes background information, while “Overview of Deep Learning Approaches for SARS-CoV-2 Analysis” furnishes a step-by-step synopsis of deep learning models. Expounding upon various deep learning techniques for SARS-CoV-2 detection, “Deep Learning Methods” offers a methodological perspective. In “Result Analysis”, a thorough exploration of recent research is presented, encompassing methodologies, approaches, data employment, outcomes, and areas of research lacunae.

Furthermore, “Discussion and Recommendation” encapsulates the analysis of results, entailing comparative and experimental assessments, ultimately culminating in the recommendation of the optimal model for COVID-19 detection. “Conclusion” is dedicated to delimitations, privacy considerations, challenges, and suggestions for forthcoming endeavours.

Rationale and intended audience of the research

Amid the SARS-CoV-2 pandemic, a thorough examination of deep learning methods for detecting the virus using X-ray and CT images is essential. Medical imaging plays a crucial role in spotting SARS-CoV-2-related complications, yet obstacles such as data scarcity and interpretability issues impede advancements. This review offers a guide for researchers and policymakers to improve diagnostic precision and patient results. By tackling present constraints and directing future investigations, it encourages innovation and cooperation in crafting efficient diagnostic instruments to combat SARS-CoV-2.

This comprehensive review is tailored for researchers and academics deeply engaged in the development and enhancement of deep learning techniques for COVID-19 detection through X-ray and CT images. Healthcare professionals, including radiologists and clinicians, seeking insights into the challenges and advancements in utilizing deep learning for diagnosing COVID-19, will find this review invaluable. Public health officials and policymakers involved in shaping strategies for COVID-19 management and pandemic response efforts can benefit from the recommendations outlined in this review. Industry stakeholders, such as medical imaging technology companies and AI developers, can gain valuable insights into collaboration opportunities and product development avenues in the field of deep learning-based COVID-19 diagnostics. Students and educators in medical imaging, machine learning, and public health disciplines will find this review a valuable educational resource to understand the current landscape and future directions of COVID-19 detection using deep learning techniques.

Article selection process

The process of article selection is elucidated in this section. Figure 2 presents a visual representation of the literature screening procedure, depicting the sequential inclusion and exclusion stages. The chosen articles stem from diverse reputable journals and were acquired through an internet database search. Electronic journals such as Scopus, IEEE Access, IEEE Explore, IEEE Transactions, Springer, ACM Library, Elsevier, SAGE, MDPI, Taylor and Francis, Wiley, Peerj, JSTOR, DBLP, and DOAJ were utilized for this purpose.

Certain articles were omitted during the study period due to specific reasons outlined in the accompanying diagram. The refinement of the final article selection transpired at both the screening and eligibility tiers. Initially, our analytical search yielded a total of 1,000 pertinent publications, from which 50 duplicate articles were eliminated. Among these articles, 750 redundant citations were excluded, leaving 200 for further examination. Applying the established criteria, 58 articles were subsequently excluded, including comprehensive reviews and review reports, while focusing exclusively on comprehensive articles that engaged in individual SARS-CoV-2 analysis.

Ultimately, the full texts of the remaining publications were thoroughly assessed, leading to the removal of 42 articles that had already been utilized in another publication addressing the same task. Consequently, our review analysis encompassed 100 manuscripts that met the stipulated inclusion criteria.

Open source data for SARS-CoV-2 analysis

This section delves into pertinent data concerning the analysis of SARS-CoV-2. Multiple common datasets are at the disposal of researchers engaged in SARS-CoV-2 detection and analysis. In this comprehensive review, our focus is on medical data, particularly two types of imaging data: CT scans and X-ray images sourced from medical repositories. These datasets originate from both public and private sources, as depicted in Fig. 4.

A succinct visualization of datasets utilized by various researchers in their SARS-CoV-2 analysis endeavours, detailing dataset types and quantities, is encapsulated in Fig. 5. This figure provides an informative overview of the dataset landscape corresponding to each model employed for SARS-CoV-2 analysis using image data. The color-coded bars denote different disease classes, while the X-axis signifies the number of data points harnessed for each model. The respective color keys for each class are conveniently provided on the right-hand side of the diagram.

Image-centric data is accessible from various repositories, with GitHub and Kaggle emerging as prime sources for freely available image datasets suitable for analysis. Predominantly, X-ray and CT scan images constitute the prevalent imaging modalities. Each of these categories possesses distinctive attributes, advantages, and limitations within the realm of research. X-ray image data offers ease of analysis and ready availability, albeit with drawbacks like the inability to detect certain conditions and manage loss of GGO density-based images. Conversely, CT scan images exhibit merits such as lower costs, heightened sensitivity, and greater reproducibility, but are accompanied by constraints such as limited accessibility and the need for multiple scans to ensure precision. Leveraging medical image processing technology could significantly contribute to SARS-CoV-2 analysis.

Furthermore, Table 2 enumerates publicly available data sources along with their corresponding links, data volumes, types, and class counts. This compilation facilitates researchers’ access to well-established and widely used datasets.

In Fig. 6, we graphically depict the count of methods subjected to analysis within our article (part a), and in part b, we present the percentage distribution of deep learning applications covered in our comprehensive review.

Several publications have presented biometric analyses of SARS-CoV-2-related experiments. Notably, the Allen Institute for AI spearheaded the CORD-19 project in collaboration with other partners, gathering a compendium of research articles centred around SARS-CoV-2 studies (Wang et al., 2020b). A significant contribution by Berkeley Lab researchers involved the development of an online search interface catering to a comprehensive compilation of scientific journal data pertaining to SARS-CoV-2. This collection encompasses diverse scholarly datasets, including the works of Wang et al. (2020b), Lite COVID, and the Elsevier Novel Coronavirus Resource Center.

Addressing the pandemic on a broader scale, the National Pandemic Initiative (NPI) amalgamates governmental efforts aimed at combatting SARS-CoV-2. To comprehensively study and interpret SARS-CoV-2 transmissions, substantial NPI datasets are requisite. Furthermore, investigations into the impact of NPI measures on SARS-CoV-2 incidents necessitate substantial volumes of data encompassing infections, fatalities, and related metrics.

In the quest for insights into SARS-CoV-2 transmission and the influence of NPI measures, a robust source of data comes in the form of the Stringency Index (Hale et al., 2020). Meticulously curated by a collective of professors and students from Oxford University, this index synthesizes openly accessible global data. Throughout the SARS-CoV-2 era, a significant portion of pertinent image data has been documented within a study journal (Shuja et al., 2021).

Data pre-processiong

Data preprocessing is an essential step prior to its practical application. The goal of preprocessing is to refine raw data into a more polished dataset. Before applying any analytical method, an initial preprocessing of the dataset is carried out to identify and rectify missing values, inconsistent data, and other irregularities. This preliminary data preprocessing stage plays a crucial role in the data mining process, involving data transformations or exclusions to ensure improved performance and reliability.

In Table 3 and Fig. 7, we provide insights into the nature of preprocessing tools utilized in SARS-CoV-2 detection from image data. These visual representations shed light on the specific preprocessing techniques employed to enhance data quality and suitability for subsequent analyses.

Types of deep learning approaches

The realm of SARS-CoV-2-related technical analysis is characterized by a multitude of methods and approaches. Within this section, we delve into the comprehensive taxonomy of SARS-CoV-2 analysis methodologies, which is vividly depicted in Fig. 8. This visual representation provides a holistic outlook on the spectrum of deep learning approaches harnessed for SARS-CoV-2 detection, analysis, and classification. The analytical pathways for SARS-CoV-2 encompass conventional techniques (statistical or manual), traditional machine learning paradigms (such as SVM (Kohmer et al., 2020; Ismael & Şengür, 2021) and random forest (Wu et al., 2020b)), as well as deep learning methods.

Central to our focus is the deep learning-based approach to SARS-CoV-2 analysis. This method is further categorized into supervised, unsupervised, and semi-supervised modalities, reflecting a comprehensive framework for harnessing the potential of deep learning more in combating the challenges posed by SARS-CoV-2.

Recently, popularly used methods are convolutional neural network (CNN) (Akl et al., 2023; Kaya & Gürsoy, 2023; Kuzinkovas & Clement, 2023; Choudhary et al., 2023; Khan et al., 2023; Sohrabi et al., 2020; Chakraborty, Murali & Mitra, 2022; Bhattacharyya et al., 2022; Islam et al., 2020), recurrent neural network (RNN) (Rayan & Alaerjan, 2023; AlMohimeed et al., 2023; Raamkumar, Tan & Wee, 2020) and Hybrid (Islam et al., 2020; Shah et al., 2021; Islam, Islam & Asraf, 2020). CNN-based SARS-CoV-2 diagnosis and prediction is the most popular. CNN method may be hybridized with attention (Ullah et al., 2023; Ouyang et al., 2020; Yang et al., 2023; Wen et al., 2023; Christina Magneta, Sundar & Thanabal, 2023), capsule, shallow, deep (Lella & Pja, 2022; Wu et al., 2020a), dynamic, deep CNN (Ardakani et al., 2020; Ozyurt, Tuncer & Subasi, 2021) and multilayer, hybrid CNN (Laguarta, Hueto & Subirana, 2020; Jain et al., 2021). In the cased of RNN (Imran et al., 2020; Hassan, Shahin & Alsabek, 2020; Imran et al., 2020), it has two evolutions named long short-term memory (LSTM) (Zhang et al., 2020) and gated recurrent unit (GRU) (Raamkumar, Tan & Wee, 2020). Both LSTM (Islam et al., 2020) and GRU (Paka et al., 2021) may be applied in bidirectional (Passricha & Aggarwal, 2020). Capsule network with RNN also gives good outcomes (capsule network) (Yuan et al., 2023; Sharma et al., 2023; Wen et al., 2023; Zhao et al., 2023; Islam et al., 2020; Prabha & Rathipriya, 2020; Malla & Alphonse, 2021; Mobiny et al., 2020; Afshar et al., 2020). Generalized adversarial network (GAN) (Bar-El et al., 2021; Fan et al., 2020b; Elzeki et al., 2021; Goel et al., 2021; Quilodrán-Casas et al., 2022; Zhu et al., 2020; Fan et al., 2020b) model works good than other method. Attention mechanism (Sitaula & Hossain, 2021; Zhang et al., 2021; Paka et al., 2021; Zhang et al., 2020; Wang et al., 2020c; Pinkas et al., 2020) sometimes obtained very good performance in SARS-CoV-2 detection. Graph conventional network (GCN) (Yu et al., 2021; Kapoor et al., 2020; Che et al., 2021; Wang et al., 2021b; Laguarta & Subirana, 2020; Liang et al., 2021) also applied on SARS-CoV-2 detection. On the other hand, ResNet (Al-Waisy et al., 2023; Islam et al., 2023a), Light CNN (Khan et al., 2023), dense and convolution (Ulukaya et al., 2023), Transfer BERT (Contreras Hernández et al., 2023; Qorib et al., 2023), RNN, multilayer perception (MLP), CNN, (Paul, Saha & Singh, 2023) etc gives good outcome in SARS-CoV-2 detection.

Deep learning algorithms can help solve complex problems by analysing fundamental depictions. Several layers are utilised progressively to learn exact interpretations as well as the property of annotated training data. Deep learning (Ardakani et al., 2020) algorithms applied in health facilities (Kong et al., 2020), smart healthcare (Esteva et al., 2019), drug development (Chen et al., 2018), medical image recognition (Kim et al., 2020), and so on. It is also widely used in the automatic diagnosis of SARS-CoV-2 victims or patients. Convolutional neural network (Ardakani et al., 2020) is applied in different transfer learning (Ahsan et al., 2021; Wu et al., 2020a). Transfer learning models are used widely those are known as such as AlexNet (Ardakani et al., 2020), GoogleNet (Ardakani et al., 2020), SqueezeNet (Ardakani et al., 2020), different versions VGG (Zebin & Rezvy, 2021; Hemdan, Shouman & Karar, 2020), diverse kinds of ResNet Wu et al. (2020b) (Wu et al., 2020b; Xu et al., 2020), Xception (Ardakani et al., 2020; Hemdan, Shouman & Karar, 2020), inception (Jain et al., 2021; Horry et al., 2020; Hemdan, Shouman & Karar, 2020), diverse types of MobileNet (Ardakani et al., 2020), DenseNet (Hemdan, Shouman & Karar, 2020), encoder-decoder (Amyar et al., 2020), capsule network (Afshar et al., 2020) etc.

Performance metrics used in COVID-19 detection

This section encompasses the key performance measures pertinent to SARS-CoV-2 analysis. Performance metrics serve as indicators of the efficacy of a given strategy. The presence of various variables and highly imbalanced datasets, stemming from diverse factors, underscores the significance of accurately gauging the annotations authenticity across the entire dataset. Consequently, these distinctive variables may not be directly comparable to conventional benchmarks.

The efficacy study seeks to demonstrate the operational capability of a system. Within the realm of SARS-CoV-2 analysis, several performance indicators are employed, including precision, recall, F-1 score, specificity, and sensitivity. These indicators facilitate the quantification of outcomes based on comparisons between anticipated and observed results. In Fig. 3, a subset of factors utilized to assess the landscape of SARS-CoV-2 research is illustrated. This depiction highlights the ongoing efforts in deep learning-based SARS-CoV-2 research, emphasizing the multifaceted evaluation process encompassing diverse aspects of performance.

Performance metrics are part of every machine learning process evaluation. Actually, it tells you if the model is developing and assigns a number to it. Deep learning techniques are indeed examples of machine learning models. Several performance metrics are used to judge the effectiveness of a deep learning model. The most common evaluation measures are listed here. P denotes positive, T indicates true, F indicates false, as well as N indicates negative in the calculation. The projected value $y_i$ in the loss function is y. The equations are used to calculate their predictions (Eqs. (1) to (5)).

(1) $$Accuracy=\frac{(TP+TN)}{TP+TN+FP+FN}$$ (2) $$Precision=\frac{(TP)}{TP+FP}$$ (3) $$Recall=\frac{(TP)}{TP+FN}$$ (4) $$F1=2\ast \frac{(Precession\ast Recall)}{Precession+Recall}$$ (5) $$Loss(y)=y_{i}log({\overline{y}}_i)+(1-y_i)+log(1-{\overline{y}}_i)$$

Basic operating step in deep learning based COVID-19 detection

Figure 9 provides an overview of the fundamental pipeline activities encompassing SARS-CoV-2 diagnosis and detection systems driven by deep neural networks. Image-based SARS-CoV-2 analysis stands as a prominent and accurate avenue of SARS-CoV-2 research. Within Fig. 9, we outline several key steps integral to image-based SARS-CoV-2 detection. Commencing with data collection from diverse sources, including open, private, and crowd-sourced origins, the participation of medical community patients is emphasized from the outset. This data, available in various formats, often employs imaging modalities such as CT scans and X-ray imaging for SARS-CoV-2 diagnosis.

A novel pre-processing methodology enhances predictive capabilities in SARS-CoV-2 analysis based on labelled image data. This involves various sub-processes like normalization, standardization, and Zero Component Analysis, all contributing to data refinement. Specifically tailored to SARS-CoV-2 imaging data comprising CT and X-ray images, the data pre-processing stage serves to eliminate undesired or extraneous elements from the images. Subsequently, pre-processed data undergoes feature analysis, encompassing feature extraction, selection, and matching. Deep learning algorithms engage in data analysis to classify SARS-CoV-2 cases following the selection of pertinent features. Within the classification phase, sub-steps encompass data partitioning, model or method selection, parameter tuning, and categorization. Data is divided into three distinct sets: training, validation, and testing, often utilizing the widely employed cross-validation technique. A model is constructed using training data, subsequently validated with the validation set, and ultimately tested using the dedicated test data. Central to the realm of computer learning-based SARS-CoV-2 diagnosis, the feature extraction and classification steps, facilitated by deep learning techniques, play a pivotal role. In this phase, deep learning automatically extracts distinctive features through iterative processes, followed by classification based on target or output classes (non-SARS-CoV-2 or SARS-CoV-2). The evaluation of the developed system’s performance encompasses metrics such as accuracy, recall, sensitivity, specificity, precision, and F1-score.

Figure 9 encapsulates the comprehensive framework of all deep learning algorithms employed for SARS-CoV-2 image analysis. At the figure’s inception, diverse forms of data are input, which are subsequently pre-processed based on their respective categories. Following pre-processing, the deep model is engaged. Deep learning models fall into several categories, including RNN, CNN, Attention, Capsule, Hybrid, and many others. Sequential implementation denotes the use of multiple levels of CNN or RNN in a consecutive sequence, which can also be layered for more intricate analysis.

CNN method

CNN is a simple image classification algorithm that is widely used. CNN is also used to classify images. LeCun et al. (1998) proposed the CNN structure for the first time (LeCun et al., 1998). Small data chunks are converted into conventional high-level vectors via a convolution stratum. The fundamental setup and corresponding CNN layout layers are shown in Fig. 10A. We also offer a brief overview of the recently built deep learning-based CNN model (Akl et al., 2023; Kaya & Gürsoy, 2023; Kuzinkovas & Clement, 2023; Choudhary et al., 2023; Khan et al., 2023) for SARS-CoV-2 analysis in the table. In classification tasks, CNN-based transfer learning methods are very common and practical. CNN-based approaches have been widely used in SARS-CoV-2 identification and classification from images in recent years. Jain et al. (2021) have developed a new method for detecting SARS-CoV-2 using a CNN classifier. This method produces a 97 percent accuracy rate on average, however, it only works with limited data and has overfitting.

Here we present the recent CNN methods of SARS-CoV-2 analysis from image (Akl et al., 2023; Kaya & Gürsoy, 2023; Kuzinkovas & Clement, 2023; Choudhary et al., 2023; Khan et al., 2023; Islam et al., 2020; Sohrabi et al., 2020; Luz et al., 2021); Fig. 11 depicts the overall process of the SARS-CoV-2 study employing all CNN-based models derived from image data. In this diagram, data is preprocessed individually before being transmitted to a CNN-based transfer learning-based system. To indicate covid as well as non-covid class, the outcome of model execution. The actual structure of each CNN model (Transfer-based) is illustrated in this diagram.

RNN method

In deep learning, RNN can handle long-ranged image features. Figure 10B depicts the basic structure of an RNN-based SARS-CoV-2 study. We also offer a brief overview of the recently built deep learning-based RNN model (Rayan & Alaerjan, 2023; AlMohimeed et al., 2023) for SARS-CoV-2 analysis in the table.

In SARS-CoV-2 case predictions utilising LSTM and RNN in GCC countries and India: Lethality and tests with several hidden layers, an LSTM model has a high degree of precision and a tiny prediction error (Razia, Arokiaraj & Jaithunbi, 2021). An RNN-based technique with CNN detected SARS-CoV-2 (Islam, Islam & Asraf, 2020) with 99.4 percent accuracy from 4,575 pictures (Islam, Islam & Asraf, 2020).

Attention method

Attention-based deep learning models (Bahdanau, Cho & Bengio, 2014) are better at handling mot relevant features than some other approaches. SARS-CoV-2 was detected with 87.49 percent accuracy using an attention-based VGG-16 classifier from 8,571 chest X-ray images (Sitaula & Hossain, 2021). This method’s data augmentation should be enhanced. SARS-CoV-2 CT images with higher precession of 94.6 percent (Zhang et al., 2021) were segmented using a dense based GAN as well as multi-layer attention model with a decent segmentation approach. The overall process diagram of attention-based SARS-CoV-2 detection is presented in Fig. 10C. We also offer a brief overview of the recently built deep learning-based attention model (Ullah et al., 2023; Ouyang et al., 2020; Yang et al., 2023; Wen et al., 2023; Christina Magneta, Sundar & Thanabal, 2023) for SARS-CoV-2 analysis in the table.

Capsule network method

The capsule and its principal capsule layer give a high level of integration for collecting as well as maintaining a complex hierarchy parallel to the various kernel dimensions (Mobiny et al., 2020). Figure 10D shows how the output is transferred to the capsule’s output layer. Overall process diagram of capsule network-based SARS-CoV-2 detection is presented in Fig. 10D. We also offer a brief overview of the recently built deep learning-based capsule model (Yuan et al., 2023; Sharma et al., 2023; Wen et al., 2023; Zhao et al., 2023; Sejuti & Islam, 2023; Basha et al., 2023; Pustokhin et al., 2023; Deepak et al., 2023) for SARS-CoV-2 analysis in the table.

A capsule network technique can be used in conjunction with RNN to improve performance. The capsule-based technique, which was recently applied in SARS-CoV-2 analysis, saves training time (Afshar et al., 2020). A capsule network might help you save time during training. SARS-CoV-2 was diagnosed with 95.7 percent accuracy from X-ray pictures using a capsule network from 13,800 images of COVIDx data (Afshar et al., 2020). A novel Radiologist SARS-CoV-2 CT scan image detection with full capsule systems with a performance of 87.6 percent accuracy (Mobiny et al., 2020). This system only works with 746 photos on a tiny domain (Mobiny et al., 2020).

Hybrid method

A hybrid deep learning-based method is constructed employing CNN, RNN, as well as other techniques, as illustrated in Fig. 10E. We offer a brief overview of the recently built deep learning-based hybrid model (Sejuti & Islam, 2023; Basha et al., 2023; Pustokhin et al., 2023; Deepak et al., 2023) for SARS-CoV-2 analysis in the table. This section describes some of the most up-to-date image classifiers required to complete the clinic intent. This section demonstrates how to analyse SARS-CoV-2 utilizing images using a recently discovered deep learning algorithm.

SARS-CoV-2 detection from chest 424 X-ray pictures with a precision of 96.00 percent (Shah et al., 2021) that used a unique Deep GRU-CNN algorithm. A deep residual network-based type attention layer with bidirectional LSTM was created for the identification and diagnosis of SARS-CoV-2. RCAL-BiLSTM is a novel ResNet-based Class Attention Layer with Bidirectional LSTM that obtains a SARS-CoV-2 diagnosis (Pustokhin et al., 2020) accuracy of 94.88 percent. SARS-CoV-2 (Islam et al., 2020) was detected using a hybrid deep learning approach with 40,000 pictures. This approach had a 99.4 percent accuracy rate. On 20,000 pictures, a deep learning-based model combining ANN to SARS-CoV-2 including feature extraction achieved 95.84 percent accuracy (Ozyurt, Tuncer & Subasi, 2021). SARS-CoV-2 was detected with 99.2 percent accuracy using a CNN-LSTM-based algorithm that classified 4,575 images (Islam, Islam & Asraf, 2020).

Comparative analysis

This section encompasses a comprehensive analysis of the overall comparison results. Undoubtedly, this segment holds significant importance within the scope of this article. Our scrutiny has been meticulously directed towards the pivotal articles in the comparative analysis, aiming to illuminate their conclusions and underscore the existing gaps in research. Inclusive within this discussion are intricate details encompassing approaches, datasets, validation procedures, and other pertinent facets.

The crux of the comparative results is encapsulated within Tables 4–8. In these tables, abbreviations are employed for succinct representation: CV signifies cross-validation, AC corresponds to accuracy, S stands for sensitivity, P denotes precision, R represents recall, T denotes training, TE represents testing, and V signifies validation. Moreover, the letters C and SP denote the number of classes and Specificity, respectively. These tables collectively provide a comprehensive visualization of the main comparative outcomes, enriching our understanding of the performance and characteristics of various methods.

Experimental analysis for model recommendation

We conducted a manual execution of 12 distinct models, meticulously tracking their individual accuracy and loss function throughout both the training and validation phases. This rigorous process was undertaken to thoroughly assess and validate the performance of newly developed models. To facilitate this evaluation, we employed the SARS-CoV-2 database, which primarily comprises X-ray images, serving as a robust foundation for comprehensively scrutinizing each model’s capabilities. The SARS-CoV-2 data used for our experimental analysis can be found at Kaggle site (https://www.kaggle.com/tawsifurrahman/COVID19-radiography-database) (Tawsif, 2022). We use 4,000 X-ray images and 1,000 photos for four image classes: normal, viral pneumonia, lung opacity, as well as SARS-CoV-2. Table 9 displays the training and validation accuracy as well as loss of twelve distinct deep learning models over 10 epochs.

Figures 12 and 13 depict graphical representation of performance (training as well as validation with loss and accuracy) for twelve models: CNN, VGG16, ResNet50, InceptionNetV3, NasNetMobile, DenseNet121, XceptionNet, AlexNet, CNN-RNN, EfficientNetB2, and MobileNetV2. Sub caption indicates the name of each model implemented. As shown in Fig. 13, loss and accuracy curves of those deep learning models offer valuable insights into their performance during training. This loss curve charts the model’s ability to minimize the error between predicted and actual outputs over time, showcasing improved learning and convergence through a declining trend. Conversely, the accuracy curve gauges the model’s success in correctly classifying inputs, illustrating its predictive accuracy. Monitoring these curves throughout training enables the assessment of model behaviour, the identification of convergence issues or overfitting, and the adjustment of training parameters to enhance overall performance. Together, these curves serve as crucial diagnostic tools for understanding and optimizing the performance of deep learning models. As shown in Figs. 12 and 13, there are some overfitting and underfitting problems in the performance of DenseNet, EfficientNet and CNN-RNN models. But other model performs well without fitting issues.

Table 9 presents the SARS-CoV-2 crucial outcome analysis. The best model is VGG-19, which has the highest training accuracy of 98.20%, while the MobileNet-v3 technique performs similarly to MobileNet, with 98.11 percent accuracy and less training loss of 0.0401. In terms of validation performance, VGG-16 outperforms the others with a validation accuracy of 94.20% and a validation loss of 0.09440. MobileNetV3 is suggested for training while VGG16 is suitable for validation activities.

Potential of the proposed method across different demographics and healthcare settings

The potential application of deep learning techniques for SARS-CoV-2 detection using X-ray and CT images across diverse demographics and healthcare settings offers significant promise. These methods can improve diagnostic accuracy for varied populations by using diverse, representative datasets. Enhancements like transfer learning and domain adaptation will boost model robustness, ensuring reliable performance with different imaging protocols and equipment. Optimizing models for low-resource environments and implementing cloud-based solutions will increase global accessibility to advanced diagnostic tools. Adhering to ethical guidelines and ensuring robust data privacy will be essential for maintaining trust and regulatory compliance.

Beyond COVID-19, these deep-learning techniques can be adapted to diagnose other respiratory diseases, aid early detection of emerging infectious diseases, and monitor disease progression. They can also be integrated into telemedicine platforms to provide remote diagnostic support, especially in underserved regions. Regular updates to the models, along with comprehensive training for healthcare providers, will keep these AI-driven solutions accurate and effective, ultimately enhancing global healthcare response capabilities and improving patient outcomes across various settings.

Limitation of deep learning model

The application of deep learning models in SARS-CoV-2 image analysis is not without its limitations, encompassing both technological and mathematical constraints. Several limitations are associated with deep learning techniques including:

• 1.

  Data quantity prerequisites for training (Subramaniam et al., 2023): Achieving significant performance gains over alternative approaches demands a substantial volume of data. Deep learning thrives on extensive datasets for effective training.

• 2.

  Computational and space complexity (Asif et al., 2023): Time and space complexity issues in SARS-CoV-2 detection using X-ray image deep learning are multifaceted. Training’s extensive duration due to complex architectures and large datasets delays model development. Deployment encounters high inference time, hampering real-time diagnosis in clinical settings. Resource-intensive computations require robust hardware, limiting deployment in resource-constrained environments. Large model and dataset sizes raise storage and infrastructure costs. Complex model interpretation hinders understanding of prediction rationale. Balancing overfitting prevention and predictive accuracy is delicate. Ensuring model generalization across diverse conditions adds complexity. Addressing these challenges is vital to harnessing deep learning’s potential for effective SARS-CoV-2 detection via X-ray images.

• 3.

  Interpretability challenges (De Falco, De Pietro & Sannino, 2023; Islam et al., 2023b): Deciphering output solely based on training is intricate, often requiring the implementation of classifiers. Techniques rooted in convolutional neural networks are commonly employed to address this challenge.

• 4.

  Black box problems (De Falco, De Pietro & Sannino, 2023): Deep learning models learn complex patterns from extensive image datasets that are not easily discernible by humans. However, their opaque decision-making process, known as the black box problem, creates trust, refinement, and application challenges, especially in clinical settings. Strategies like interpretable machine learning, model distillation, adversarial training, self-explaining neural networks, and visualization techniques aim to address this issue. Ongoing research seeks to enhance deep learning model reliability and usefulness, particularly in clinical contexts. The black box problem is particularly significant in deep learning due to its reliance on concealed patterns in large image datasets. This opacity undermines confidence in predictions, vital in healthcare where understanding model operations is key for informed decisions. Debugging and improving models are complicated by this problem, as identifying issues without comprehending internal workings hampers performance enhancements.

• 5.

  Overfitting problems (Meedeniya et al., 2022): Overfitting is a concern in machine learning, including deep learning for SARS-CoV-2 detection via X-ray images. It happens when a model becomes overly attuned to training data, hindering its ability to generalize to new data. Consequently, this can result in inaccurate predictions. Overfitting arises from factors like a small training dataset, limiting the model’s exposure to genuine patterns. Excessive model complexity exacerbates this issue by causing the model to memorize data rather than learn meaningful insights. Regularization, a technique to prevent overfitting, introduces a penalty to the model’s loss function, promoting simplicity and curbing the model’s tendency to grasp irrelevant training data patterns.

• 6.

  Lack of contextual understanding (Hu et al., 2022): Deep learning models used for SARS-CoV-2 detection from X-ray images learn patterns associated with the disease but lack contextual understanding, leading to potential inaccuracies in predictions and challenges in explanation. For instance, the model might identify SARS-CoV-2-related features without comprehending their significance within the overall image. This hinders its ability to explain predictions and poses trust issues in clinical settings. To tackle this limitation multimodal data, transfer learning methods and interpretable machine learning can be employed. Real-time SARS-CoV-2 data problems:

• 7.

  Real-life data problems (Srivastava et al., 2022): Real time data helps to detect and predict disease precisely. In SARS-CoV-2 detection using X-ray image deep learning includes the following real time data problems:

• (a)

  Data scarcity (Ali, Grönlund & Shah, 2023): Limited SARS-CoV-2 data due to the disease’s novelty hampers accurate model training, reducing comparability with diseases possessing more data; (b) Data imbalance (Calderon-Ramirez et al., 2021; Chamseddine et al., 2022): Unequal representation of SARS-CoV-2 cases vs non-SARS-CoV-2 cases makes it challenging for models to correctly identify the disease; (c) Data Noise (Momeny et al., 2021): Noisy data containing artifacts interferes with model learning, impacting accuracy; (d) Data bias (Catalá et al., 2021): Biased data not representing diverse populations reduces model accuracy for certain groups.

These drawbacks compound other issues such as complex model training, lack of robustness to real-world variations, and computational resource demands during deployment. Despite these hurdles, addressing these problems through research efforts can lead to improved deep-learning models for early SARS-CoV-2 detection and diagnosis using X-ray images.

Ethical considerations and privacy issues in SARS-CoV-2 analysis

Ethical concerns and validation are paramount when using patient images for AI training in COVID-19 detection. Ensuring informed consent from patients is essential, alongside robust anonymization procedures to safeguard privacy. Compliance with regulations like General Data Protection Regulation (GDPR) (Varpula et al., 2024; Tertulino, Antunes & Morais, 2024) and Health Insurance Portability and Accountability Act (HIPAA) (Carmine, 2024) is mandatory to ensure data protection. Validation of AI models must involve rigorous testing to ensure accuracy and reliability, with transparency in the validation process to maintain trust. Addressing potential biases in the training data is crucial to ensure fairness and equitable outcomes (Vrudhula et al., 2024). Ultimately, adherence to ethical guidelines and validation protocols is essential to uphold patient rights and ensure the ethical use of AI in COVID-19 detection (Abedi et al., 2024).

Preserving privacy in the context of SARS-CoV-2 data analysis remains a critical concern. The encryption and secure management of medical images related to SARS-CoV-2 has been suggested as an effective approach (Durafe & Patidar, 2022). Within this realm, the emphasis is placed on ensuring the confidentiality, accessibility, and security of SARS-CoV-2-related information (Shuja et al., 2021). Big data challenges intertwined with SARS-CoV-2 monitoring methods encompass issues of security, safety, and sustainability, especially pertinent in the context of pandemic-induced health concerns. Managing large datasets within data centres during SARS-CoV-2 has posed challenges, particularly in the areas of AI, deep learning, and statistical research that utilize user data from sources like medical centres and social media. The omnipresent concern for privacy continues to be a focal point, with the existence of open-source data exacerbating privacy issues. These privacy concerns can sometimes overshadow critical public health matters, potentially leading to constraints on the sharing of scientific data (Shuja et al., 2021).

There is a looming apprehension that post-pandemic, surveillance practices may persist beyond the realm of public health. SARS-CoV-2-related applications have facilitated data exchange with third parties, raising legitimate concerns about user privacy (Naudé, 2020). While patient data collection has been supported by hospitals and medical institutions, concerns persist regarding the thorough assessment of stability and privacy vulnerabilities within these datasets, despite efforts towards data anonymization. The automatic contact tracing measures adopted by governments in response to SARS-CoV-2 transmission have also prompted deliberations over consumer privacy considerations (Chan et al., 2020).

In the pursuit of scholarly purposes, data sharing mandates meticulous protection of both physician and mobility data to avert biases. Federated deep learning (DL)-based learning approaches offer a potential solution, as they circumvent the need to centralize data at a single cloud data centre. In federated DL, model parameters and outputs are communicated between participating nodes, ensuring data protection within a cohesive machine learning framework (Liu et al., 2020). Furthermore, in the context of the SARS-CoV-2 pandemic, the prioritization of public health concerns may take precedence over matters of individual privacy (Latif et al., 2020).

In summary, the intricate balance between data utilization for vital research and safeguarding privacy remains a central challenge in SARS-CoV-2-related data analysis.

Challenges in SARS-CoV-2 analysis

Challenges of different types of deep learning methods are addressed here, the challenges can boost researchers to develop a new and improved method to analyse SARS-CoV-2.

1. SARS-CoV-2 data generation (Phukan et al., 2022): Manual data generation is challenging in the field of health-related concern. We should be assisted by health or clinical institutions to get patient data. Permission and government allowance to make available for analysis is the big challenging issues in data generation.

2. Data availability (Tsai et al., 2021): Most of data set for SARS-CoV-2 analysis is public and some are private. To analyse all these data sets, we have to be aware about its verification of labelling, augmentation and balancing (Tsai et al., 2021).

3. SARS-CoV-2 data labelling legislation and verification (Yang et al., 2021): SARS-CoV-2 results are exchanged according to strict regulations in countries all over the world. One of the primary procedures specifically states that the bare minimal amount of specimens and data from infected individuals must be gathered in the shortest amount of time. As a result, analysis is becoming more and more difficult to perform. Before using data to analyse, proper data labelling and its correctness need to be confirmed by the source and its released licence.

4. SARS-CoV-2 data augmentation (Nishio et al., 2020): SARS-CoV-2 data augmentation approaches are becoming more popular for analysing output quality. The data augmentation field demands new research and study in order to provide fresh/synthetic data with sophisticated applications. Artificial Intelligence tough to use GANs to create high-resolution images, for example. If there are biases in the original dataset, there will be biases in the data supplemented from it. As a result, figuring out the optimal data augmentation strategy is essential.

5. Noisy SARS-CoV-2 data (Momeny et al., 2021): The presence of noisy SARS-CoV-2 data in a data set can have a major impact on the accuracy of any useful information prediction. Much empirical research has demonstrated that data set noise significantly reduced classification accuracy and resulted in poor prediction outcomes.

6. Quality and dimension of SARS-CoV-2 image data (Rajinikanth et al., 2020): SARS-CoV-2 image quality should be good quality other it may lead deep learning classifier to predict wrong classification. Normally, most of the raw data from medical institutes are in 3D form and deep learning models are not suitable to be trained with 3D data, it can deal with 2D data. So it should be a concern and solved.

7. Imbalanced SARS-CoV-2 data (Calderon-Ramirez et al., 2021; Chamseddine et al., 2022): Because of the significantly lopsided class distribution and disproportionate misclassification costs, balanced categorization is particularly difficult. Properties including dataset size, label noise, and data dispersion add to the difficulty of imbalanced classification. Resampling SARS-CoV-2 data and other tools can handle imbalanced data.

8. SARS-CoV-2 data preprocessing (Maity, Nair & Chandra, 2020): There are some challenges in SARS-CoV-2 dat preprocessing like: missing data, manual input, data inconsistency, regional formats, wrong data types, file manipulation, and missing anonymization. Data preprocessing with data cleaning, data integration, data transformation, and data reduction are all examples of data preprocessing. Data cleaning is a technique for removing noise and correcting data discrepancies.

9. Model development, improvement (Chauhan, Palivela & Tiwari, 2021) and tuning (Lee et al., 2020) for SARS-CoV-2 detection: Because of the significantly lopsided SARS-CoV-2 class distribution and disproportionate misclassification costs, balanced categorization is particularly difficult. Properties including dataset size, label noise, and data dispersion add to the difficulty of imbalanced classification. Deep learning models have challenges if there are the following issues like: not enough training data, poor quality of data, irrelevant features, non-representative training data, overfitting and underfitting.

10. Intensive training expenses (Gupta et al., 2023): The complexity of SARS-CoV-2 data models results in resource-intensive training processes. The computational requirements necessitate the use of numerous workstations and costly GPUs, leading to elevated operational expenses.

11. Skill and expertise demand (Attallah, 2023): Deep learning lacks a standardized model selection process. Optimal selection demands an understanding of topology, training methodologies, and other intricate characteristics, often posing challenges for individuals with limited expertise.

Future scopes

A plethora of deep learning model implementations has been applied across diverse datasets, with a particular emphasis on radiology imaging datasets, as showcased and explored in prior research (Liu et al., 2020; Nishio et al., 2020). However, the translation of these techniques into real-world patient care scenarios remains challenging, underscoring the pressing necessity for benchmarking frameworks that can assess and compare novel methodologies. Such frameworks can facilitate the practical deployment of analytical hardware infrastructures by accounting for common medical data, data preprocessing procedures, and assessment requirements across various AI approaches, thereby ensuring data transparency and interpretability (Tsai et al., 2021; Phukan et al., 2022).

In today’s world, advancements in sophisticated technologies like artificial intelligence are pervasive and essential in combating SARS-CoV-2. Integrating AI can not only offer intrinsic advantages but also mitigate misinformation dissemination on social media platforms (Bharati et al., 2021). By analyzing vast amounts of data, AI ensures the dissemination of accurate information, preempting premature panic and hysteria amidst the pandemic.

Cross-disciplinary collaboration is paramount in advancing the field of COVID-19 detection, as it enables researchers to gain insights into clinical needs and challenges from diverse perspectives (Ritto et al., 2024). This collaboration between computer scientists, medical professionals, and public health experts ensures the development of AI models that will be effective and clinically relevant. Moreover, the development of accessible models is crucial for widespread adoption in healthcare settings. Future research should prioritize user-friendly interfaces, comprehensive training programs, and compatibility with existing healthcare systems to facilitate seamless integration. Additionally, achieving generalizability across different demographics, regions, and healthcare settings is essential (Abad, Casas-Roma & Prados, 2024). Transfer learning techniques and domain adaptation methods can enhance model performance and adaptability. Emphasizing the interpretability of AI models is vital for building trust among healthcare professionals. Interpretable models provide clear explanations for predictions, enabling clinicians to understand and trust the AI-driven diagnostic process (de Sousa Freire et al., 2024). Furthermore, real-time detection techniques should be explored to enable rapid diagnosis and treatment decisions. Optimizing model architectures for speed and efficiency is crucial for achieving real-time performance (Kaur et al., 2024). Models should also be robust to variations in imaging quality, patient demographics, and disease severity to maintain diagnostic accuracy in real-world scenarios (Khero, Usman & Fong, 2024).

The integration of various types of medical and serological clinical data significantly enhances the diagnostic accuracy and robustness of models for COVID-19 detection. By providing a more comprehensive view of the patient’s health, this approach helps identify the virus more reliably, thereby reducing the likelihood of false positives and negatives. This comprehensive integration ultimately supports more precise and effective patient care (Lei & Mohan, 2020; Kamari et al., 2024).

Ethical considerations are paramount in the development and deployment of AI models to ensure patient privacy, informed consent, and equity in healthcare delivery. By integrating ethical practices, researchers can contribute to the responsible use of technology in healthcare and mitigate potential biases or unintended consequences (Bartenschlager et al., 2024).