A scoping review of literature on deep learning and symbolic AI-based framework for detecting Covid-19 using computerized tomography scans

This scoping review aims to explore various Deep Learning and Symbolic Artificial Intelligence (AI) models that can be integrated into explainable hybrid AI for the purpose of detecting COVID-19 based on Computerized Tomography (CT) scans. We followed the PRISMA-ScR framework as the foundation for our scoping review protocol. Our approach included a thorough search across 13 databases, complemented by an additional random internet search for relevant articles. Due to the voluminous number of articles returned, the search was further narrowed using the keywords: Deep Learning, Symbolic AI and Hybrid AI. These keywords were used because they are more visible in the earmarked literature. A screening of all articles by title was performed to remove duplicates. The final screening process centered on the publication year, ensuring that all considered articles fell within the range of 2019 to 2023, inclusive. Subsequently, abstract or text synthesis was conducted. Our search query retrieved a total of 3,312 potential articles from the thirteen databases, and an additional 12 articles from a random internet search, resulting in a cumulative count of 3,324 identified articles. After the deduplication and screening steps, 260 articles met our inclusion criteria. These articles were categorized based on the year of publication, the type of aim, and the type of AI used. An analysis of the year of publication revealed a linear trend, indicating growth in the field of Hybrid AI. Out of the five aim categories identified, we deliberately excluded articles that lacked a specified aim. It's noteworthy that 3% of the articles focused on the integration of AI models. The low percentage value suggests that the integration aspect is overlooked, thereby transcripting the integration of Deep Learning and symbolic AI into hybrid AI as an area worth exploring. This scoping review gives an overview of how a Deep Learning and Symbolic AI-based framework has been used in the detection of COVID-19 based on CT scans. © 2024 by the authors. Licensee SSBFNET, Istanbul, Turkey. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).


Introduction
Deep Learning algorithms have showcased their potential and solidified their position as prominent AI techniques, particularly in the field of chest radiography, for detecting anomalies, irregularities, and diagnostic patterns (Amyar et al., 2020).In the same vein, Deep Learning algorithms have made notable progress in areas such as object identification (Wu et al., 2020), semantic segmentation (Liu., 2019), and image classification within the realm of image recognition tasks (Chai et al., 2021).This has elevated Deep Learning algorithms to the standard gold for the binary classification problem, particularly binary classification of CT scan images for COVID-19 diagnosis (Wang et al., 2021).Deep Learning represents an AI extension that employs dense layers for entirely automated processes of feature extraction, classification, and segmentation (Ker et al., 2017;Razzak et al., 2018).It relies on Artificial Neural Networks with an emphasis on representation learning (LeCun et al., 2015).Deep Learning algorithms empower computational models, which consist of multiple processing layers, to acquire data representation across various levels of abstraction (Alzubaidi et al., 2021).They train a computer model to directly undertake classification tasks using images, text, or audio data, achieving a high level of accuracy in the process (LeCun et al., 2015).Nonetheless, the computational process of Deep Learning models remains opaque to human understanding, and as a result, these Deep Learning models are commonly referred to as 'black boxes' (Saleem et al., 2022).The decision-making mechanisms of these neural networks cannot be elucidated, raising significant concerns regarding the trustworthiness and transparency of Deep Learning models (Beaudouin et al., 2020).The absence of explainability in Deep Learning algorithms hampers researchers and model developers in recognizing potential shortcomings and opportunities for enhancement (Giuste et al., 2022).Cutting-edge methods might be discovering suboptimal solutions, or, in a more unfavorable scenario, constructing their solutions using irrelevant input features (Su et al., 2019;Eykholt et al., 2018).Is it feasible to employ Deep Learning models, particularly in healthcare and safety-critical scenarios, when we lack a comprehensive understanding of their internal processes, especially considering that even a slight chance of an erroneous decision could potentially endanger human lives?
In contrast, Symbolic AI is focused on the explicit representation of human knowledge in a declarative format (Singh, 2019).Symbolic AI entails the deliberate integration of human knowledge and behavioral rules into computer programs (Dickson, 2019).Essentially, Symbolic AI employs human-readable symbols to symbolize real-world entities or concepts, incorporating logic to establish rules for the precise manipulation of these symbols, resulting in the creation of a rules-based system (Verani, 2020).Symbolic reasoning involves the manual creation of rules through human intervention, which are subsequently hardcoded into a fixed program.In contrast, in Deep Learning, the algorithm learns rules by establishing correlations between inputs and outputs (Gill, 2023).When employing Symbolic AI, everything is transparent, comprehensible, and explainable, resulting in what is commonly referred to as a 'transparent box,' in contrast to the 'black box' approach used in Deep Learning (Verani, 2020).Given that it operates on a rules-based system, a Symbolic AI program can readily provide an explanation for why a particular conclusion is reached and outline the reasoning steps involved (Younis, 2022).It's important to highlight that the drawbacks of Deep Learning align with the strengths of Symbolic AI, and conversely, the weaknesses of Symbolic AI complement the strengths of Deep Learning (Gijo, 2023).So how can we exploit the complementary nature of Deep Learning and Symbolic AI?An essential principle within the symbolic paradigm is that intelligence emerges from the manipulation of abstract compositional representations, where the elements symbolize objects and their relationships (Garnelo and Shanahan, 2019).Hence, a primary aim of Deep Learning is to create architectures that can identify objects and relationships within unprocessed data, learning effective representations of them for subsequent processing (Garnelo and Shanahan, 2019).Therefore, the need for the integration of Deep Learning and Symbolic AI becomes apparent.
The integration of Deep Learning and Symbolic AI towards Hybrid AI approaches has not been receiving adequate attention despite its promising potential in a number of research areas including medicine, bioinformatics, visual intelligence and ontology learning (Gupta, 2021).At the core of Hybrid AI models is the ability to leverage the strengths of Deep Learning and Symbolic AI while remaining explainable (Thomas, 2022;Lawson et al., 2021).The sudden emergence of Coronavirus Disease 2019 (COVID-19) brought an opportunity to exploit this combinatorial power of Deep Learning and Symbolic AI towards highly accurate and explainable hybrid AI models to detect the disease (Huang et al., 2021).The integration aspect, therefore, presents an area worth examining.This also signifies that the integrated neural-symbolic systems theory has reached a level of maturity that warrants practical testing with real-world application data (Bhatia 2017).
This article aims to comprehend the breadth of hybrid AI literature that discusses the crucial Deep Learning and Symbolic AI features in order to deduce solutions to the COVID-19 challenge.Our objective is to pinpoint and categorize the diverse literature that delves into Deep Learning, Symbolic AI and Hybrid AI.This involves scrutinizing the categories of emphasis, research contexts, methodologies employed, key variables, parameters of interest, and the salient attributes of associated models.It is envisaged that this would result in explainable Hybrid AI relevant to the COVID-19 problem.

Research and Methodology
It is crucial to elucidate how the PCC framework influenced the scoping review's process of selecting the appropriate articles for the study at hand.Furthermore, it is advisable to delve into the sources of the articles included in this scoping review and comprehend the criteria that guide the screening of articles.This section provides an in-depth exploration of these aspects before we outline the reporting procedure that summarizes the process of evidence synthesis.

The PCC Framework
The Joanna Briggs Institute (JBI) reviewer's guidebook presents a range of mnemonic devices tailored to various types of reviews, with the PCC framework being one such mnemonic implied in reference (Peters et al., 2015).In the methodology employed for this scoping review, we harnessed the PCC framework to identify the relevant articles within the domain of Hybrid AI and their application in detecting COVID-19 using CT scans.The guiding question for shaping the crucial concepts and contextual factors that informed our search criteria and article inclusion criteria was as follows: "What literature on Deep Learning, Symbolic AI and Hybrid AI, where, when, why and how, has

it been deduced in detection of COVID-19 based on CT scans? What apparent gaps exist?"
To tackle the challenge of COVID-19 detection, we cast a wide net, considering a broad spectrum of articles related to Deep Learning, Symbolic AI, and Hybrid AI.As a result, our primary focus revolved around the concepts of Deep Learning, Symbolic AI, and the elucidation of Hybrid AI in the context of COVID-19 detection.Our overarching goal throughout this study was to pinpoint facets of Deep Learning and Symbolic AI that could be incorporated into the development of Hybrid AI ontologies.

Sources of Articles
We conducted an extensive literature search across thirteen online databases to assemble the pool of pertinent articles for this scoping review.These databases encompassed JSTOR, ACM Digital Library, Cambridge Core, Springer, ProQuest, EBSCOhost, IEEE Xplore Digital Library, Emerald, ERIC, GreenFile, ScienceDirect, MasterFile, Premier Scopus, and Taylor and Francis.We considered articles published from 2019 to 2023, inclusively, to ensure the relevance and currency of the literature under scrutiny.
We utilized the JBI scoping review protocol (Peters et al 2015;Bpharm et al., 2020) to select articles for our study, focusing on those that encompassed one or more of the following key concepts: Deep Learning, Symbolic AI, explainable Hybrid AI, and the detection of COVID-19 using CT scans.Priority was given to papers that centered on the practical application of Deep Learning, Symbolic AI, and explainable Hybrid AI in the context of COVID-19 detection via CT scans.We limited our selection to peer-reviewed conference papers and journal publications in English, without specific regard to the depth of the articles' core emphasis.A detailed analysis of the articles extracted from these databases will be conducted in Section 3.

Search Strategy, inclusion and exclusion criteria
We developed the LitSearchTool(), a tool for configuring literature settings, with the aim of facilitating an objective, uniform, and replicable approach to text mining from the databases.The LitSearchTool() possesses the capability to identify synonymous text patterns that might be missed by human readers.Moreover, it mitigated the potential bias that could arise when individuals select keywords based on their personal interpretations.The use of the LitSearchTool() promoted uniformity and reproducibility in the various stages of the literature search procedure.The following code snippet provides a concise representation of the query initiated and executed through the LitSearchTool().

DB<GreenFile\ProQuest\Emerald\ERIC\CambridgeCore\IEEEXplore\JSTOR\
MasterFilePremierEBSCOhost\ScienceDirect\Scopus\Taylor & Francis> We adopted a sequential approach for database selection, with each population of relevant studies comprising articles that encapsulated the key concepts within their appropriate context.To further validate the chosen set of articles, our standard procedure involved manual screening, encompassing titles, abstracts, introductions, and, in some cases, complete articles.To streamline this process, we explored the possibility of leveraging the revtools R package to identify and eliminate duplicates.Additionally, we employed the tool for screening article titles and abstracts.Our utilization of the revtools package was guided by the PRISMA-ScR framework (Tricco et al., 2020), which facilitated the creation of a PRISMA flow diagram illustrating the screening process.This diagram delineated exclusions at each step and highlighted the final selection of eligible articles.

Data extraction and collation
Articles that satisfied the inclusion criteria were abstracted to include information such as the publication year, their primary objectives, the specific AI methodologies employed, and how these AI techniques were applied in the context of COVID-19 detection.The synthesis and representation of knowledge extracted from eligible articles, as well as the areas of research yet to be explored, were predominantly conveyed through the use of figures, charts, and tables.

Results
In this section, we provide an overview of our findings through a PRISMA-ScR diagram.Following this, we delve into the various categories of insights generated by our analysis.Notably, we identify areas in the existing literature where gaps exist, which may serve as valuable prompts for future studies within the framework of Deep Learning and Symbolic AI for COVID-19 detection using CT scans.

The PRISMA-ScR
Figure 1 illustrates the PRISMA-ScR diagram, offering a visual representation of the article screening workflow we followed.In addition to our systematic search across the thirteen selected databases, we conducted an additional internet search using the same search query and criteria outlined in Section 2.3.This extra step was taken to ensure comprehensive coverage and minimize the risk of overlooking any potentially eligible articles.As a result, we identified a total of 3,312 articles that adhered to the PCC framework.
To assist in the article screening process, we employed a SQL query in conjunction with the revtools software.During the initial screening phase, we identified and removed a total of 531 articles as duplicates.These duplicates shared similarities such as identical titles, authors, publication years, and subject matter.Listing 1 presents the SQL query used in the deduplication procedure.It is worth noting that some duplicate articles may have arisen due to their presence in multiple web databases.
select * from ( select row_number () over (partition by Title, Author, Year), title, author, year from ScopingReview) where t.row_number < 2; Listing 1: Query extracted from revtools R package Subsequently, for the remaining articles with similar-sounding titles, we conducted a second round of screening involving a thorough examination of their content.Using the same revtools R package, we identified 11 articles with titles that exhibited marginal differences, despite their content essentially presenting the same study.As a result, we further refined the selection by eliminating six additional publications through a textual synthesis process.In these cases, even though the articles had different titles, their abstracts led to essentially the same research.Furthermore, we excluded the last 203 articles because they were published before the year 2019.Following the completion of this second round of screening, a total of 2,563 articles were determined to be both valid and relevant.
During the third and final round of screening, a significant portion of the initial population was eliminated.In this phase, articles were excluded if they did not align with the overarching context of utilizing Deep Learning, Symbolic AI, or Hybrid AI for the purpose of detecting COVID-19 through CT scans.Additionally, some articles were rejected because the techniques they employed did not consistently fall within the domains of Deep Learning, Symbolic AI, or Hybrid AI.As a result of this rigorous round of screening, a total of 260 articles remained eligible and were subsequently considered for further analysis in the subsequent phases of this scoping review.

Categories of eligible articles
Within the framework of a Deep Learning and Symbolic AI-based approach for detecting COVID-19 through CT scans, the 260 papers that met the criteria for this review were subjected to a comprehensive analysis and categorized into four distinct groups.The first category involved an examination of the distribution of these articles based on their year of publication.We also paid close attention to the primary objectives and areas of focus addressed in these papers.Furthermore, we scrutinized the classification of these publications according to the specific type of AI each qualifying article represented.Lastly, we grouped the included articles based on the AI techniques employed in the detection of COVID-19 through CT scans.The subsequent part of this section provides a detailed overview of the findings within each of these categories.

Distribution of the included articles by year of publication
As part of our inclusion criteria, we focused on articles published between 2019 and 2023.Analyzing the distribution of the 260 included publications by year allows us to observe trends in the growth of research related to Hybrid AI in the context of detecting COVID-19 through CT scans.The distribution of the publications over the years is visually presented in Figure 2. The generated frequency distribution can be fitted with a linear trend line, indicating a consistent and steady increase in research within this domain.Notably, there is a significant drop in the number of articles published in 2023, possibly due to some articles still being under review for publication at the time of concluding this study.This observed upsurge in researchers' interest in exploring this subject highlights the relevance of our study and confirms the genuine effort to address a prevalent research agenda.Our aim to contribute to this evolving field of study, specifically in the understanding of Hybrid AI systems for detecting COVID-19 via CT scans, represents a promising avenue for further exploration in future studies.

Distribution of the included articles by aim
In the context of this study, an 'aim' refers to the underlying intention, goal, focal point, or objective that propels the research.During the scoping review, it became evident that most AI studies exhibited five overarching types of goals, which could be categorized into the following groups: (a) investigation, (b) review, (c) comparison, (d) evaluation, and (e) integration of discrete models.However, there were instances where the objectives and aims were implicit and challenging to neatly classify.Figure 3 illustrates the distribution of the identified research goals among the included articles.Most of the articles exhibited a lack of clarity regarding their specific objectives.Nonetheless, the broader analysis underscores the prevalence of articles that emphasize the investigation of AI models.Conversely, only a limited number of articles concentrated on the integration approach, indicating that this aspect is often overlooked.Therefore, the integration facet appears to be an underexplored area warranting further investigation.

Distribution of the included articles by type of AI
Artificial Intelligence involves machines, particularly computer systems, simulating human intelligence processes.Our primary emphasis has been on Deep Learning, Symbolic AI and explainable Hybrid AI in the context of detecting COVID-19 through CT scans.The objective is to assess the prevalence of each AI approach in the detection of COVID-19 and gain valuable insights into their effectiveness and rationale for solving this problem.Table 1 provides   The following section provides a high-level overview of the primary observations derived from the study, which will be followed by our conclusions, contributions, and a glimpse into potential directions for future research.

Conclusion
We draw several noteworthy insights from the study's findings.The need for advancement in the domain of a Deep Learning and Symbolic AI-based framework for COVID-19 detection via CT scans is clearly evident through the observed growth in attention over the years.The integration of Deep Learning architectures with Symbolic AI holds promise as a promising direction.It's worth emphasizing that Hybrid AI harnesses the strengths of both Deep Learning and Symbolic AI to construct an interpretable model, a trend gaining traction, particularly in the healthcare sector where a small margin of error in decision-making could have severe consequences for human life.
This study brings to light three key observations, which are as follows: i.
A key question is unanswered.In this case, which key component units of Deep Learning models and Symbolic AI models can be recommended towards the design of an explainable Hybrid AI model for the detection of COVID-19 based on CT scans?The upcoming study will consider key component units which are likely to improve the accuracy and explainability of the resultant Hybrid AI. ii.
We also observe that explainability is a key feature of the desired model.How do we design and integrate Deep Learning and Symbolic AI views into an explain-able model?In responding to this question, it is crucial to consider how to incorporate pertinent Symbolic AI concepts into a Deep Learning framework to create a useful hybrid.iii.
Achieving a high level of accuracy is at the core of AI models.To what extent does the proposed Hybrid AI model's accuracy compare to the accuracies of stand-alone Deep Learning and Symbolic AI models in successfully detecting COVID-19 based CT scans?
The evident contributions of this scoping review can be summarized in two aspects: i.
The scoping review revealed conspicuous disparities in the distribution of articles within the Deep Learning and Symbolic AI-based framework for detecting COVID-19 via CT scans.These imbalances are evident both in the stated objectives of the articles and the preference for Hybrid AI technologies.These imbalances are evident both in the stated objectives of the articles and the preference for Hybrid AI technologies.This insight opens up opportunities for more comprehensive investigations and in-depth research within this domain.Our scoping review has established a foundational pathway upon which future innovative research endeavors can be constructed.ii.
This study has offered valuable educational research perspectives in the field, which newcomers to this domain can leverage.This scoping review can serve as an excellent initial reference for addressing these perspectives.
Subsequent research efforts should be directed toward combining connectionism and symbolic approaches within the field of AI, with the aim of unlocking new opportunities to develop intelligent systems capable of decision-making.The focus should center on establishing integrated neural-symbolic systems, with the ultimate goal of achieving more robust reasoning and learning systems for applications in computer science.

Figure 2 :
Figure 2: Included articles distribution by year of publication

Figure 3 :
Figure 3: Distribution of Included Articles by Aim an overview of the prevailing utilization of various AI approaches in the included articles, which can be broadly classified into four categories.In certain cases, conventional machine learning methods were employed either in conjunction with Deep Learning or independently to constitute Hybrid AI, leading to enhanced model performance.The remarkable predominance of Deep Learning and Hybrid Deep Learning in the realm of COVID-19 detection through CT scans is quite striking.This is likely attributed to their consistently high levels of performance and accuracy.However, it raises a pertinent question: Do Deep Learning and Hybrid Deep Learning always yield interpretable and explainable decisions?Conversely, Hybrid AI models are renowned for delivering interpretable and highly accurate outcomes.This sparks inquiries like: (a) What attributes of Hybrid AI contribute to explainability and performance enhancement?(b) What specific elements are absent in other AI systems but present in Hybrid AI models?Notably, this scoping review highlights the potential for further research to delve into these inquiries from diverse perspectives.It may also be valuable to explore the critical factors within Hybrid AI models that underlie improved accuracy.Additionally, experimental investigations could ascertain whether achieving the outcomes of Hybrid AI models is attainable with AI other than Deep Learning and Symbolic AI.

Table 1 :
Distribution of included articles by type of AI