The evolution of Deep Learning (DL) methodologies has significantly enhanced the field of medical Imaging, particularly in the interpretation of chest radiographs (CXRs). Among these advancements, Convolutional Neural Networks (CNNs) have emerged as a paramount technology for processing and Classifying CXR images and demonstrating exceptional proficiency in detecting COVID-19-related signs (Perumal et al., 2020). Although Reverse Transcription Polymerase Chain Reaction (RT-PCR) tests surpass CXRs in accuracy and reliability for virus detection, the latter remains an ubiquitous tool in clinical practice (Giri and Rana, 2020). RT-PCR excels in early detection capabilities, enabling prompt treatment initiation, and uniquely identifies the virus in asymptomatic individuals through the analysis of: Saliva samples, throat samples, and nasal passage samples, demonstrating superior performance over CXR evaluations in these aspects.

A recent investigation employed a Convolutional Neural Network (CNN) model to distinguish between:

The CNN model, trained on an extensive dataset comprising images from:

Healthy subjects achieved a remarkable precision rate of 97.6% in identifying COVID-19 cases on chest radiographs (CXR). This exceptional level of precision surpasses that of traditional CXR diagnostic techniques, demonstrating the potential of deep learning algorithms to improve diagnostic accuracy (Wang and Lin, 2020). In the current healthcare landscape, the surge in demand for intensive care units (ICUs) has exposed the capacity constraints of healthcare systems in several developed countries. The influx of patients suffering from COVID-19-induced pneumonia into the ICU underscores this pressing challenge, highlighting the need for effective diagnostic tools and strategies to manage the burden on healthcare resources (Kermany et al., 2018).

The system employs relational feature intelligence to analyze and interpret the interactions among various elements within an image, enabling the assessment of:

This capability facilitates the evaluation of:

The relational feature intelligence of the system enables the analysis of complex interactions within medical images, providing valuable information on spatial relationships and dynamics between different elements (Ai et al., 2020; Ng et al., 2020; Kong and Agarwal, 2020). This capability is particularly beneficial in:

As depicted in Figure 1, the radiographic features characteristics of COVID-19 typically encompass:

In contrast, viral pneumonia caused by non-SARS-CoV-2 viruses, such as influenza-A, tends to exhibit:

as a primary diagnostic strategy for early identification of this novel pneumonia strain.

The primary objective of this examination is to detect the presence of the virus within a patient, which is achieved through the analysis of specific biomarkers (Choe et al., 2019). To augment this assessment, medical practitioners often employ a multimodal diagnostic approach, incorporating additional methodologies such as:

to verify infection and facilitate accurate diagnosis. However, it is crucial to acknowledge that these supplementary tests may not always yield precise outcomes, highlighting the necessity for their combined application with other diagnostic techniques to ensure a definitive diagnosis.

A significant limitation of this specific examination is the considerable:

associated with its execution, which may hinder its applicability in resource-constrained settings.

To minimize diagnostic errors, healthcare professionals are advised to utilize automated imaging analysis software, powered by artificial intelligence (AI), when interpreting X-ray photographs. These advanced tools excel at identifying complex patterns within images, thereby improving diagnostic accuracy. Additionally, to maintain a high level of diagnostic precision, medical staff should take periodic breaks to avoid fatigue and seek second opinions from colleagues to ensure reliability and accuracy. Innovations in diagnostic methodologies, such as the integration of AI and machine learning (ML) technologies, can significantly reduce both the time and financial costs associated with diagnostic processes, enhancing efficiency and reliability. Specifically, AI and ML can accelerate and refine disease diagnosis by analyzing medical data. In addition, digital biomarkers, such as data from wearable technologies, offer real-time information, enabling the early detection and monitoring of various health conditions. The synergy of ML and AI has facilitated the development of sophisticated algorithms capable of detecting and diagnosing diseases with remarkable precision. These advancements have transformed the diagnostic landscape, enabling healthcare professionals to make more accurate and timely diagnoses (Simonyan and Zisserman, 2015). Artificial intelligence (AI)-empowered systems can efficiently process vast datasets, enhancing the capacity for early disease detection and the formulation of customized treatment plans. The advent of telemedicine and remote monitoring tools enables healthcare practitioners to oversee patient health remotely, providing timely interventions when necessary. This approach has been particularly advantageous during the pandemic. Pre-trained models can address complex issues with speed and accuracy, offering a significant advantage in medical diagnostics. Leveraging pre-trained models reduces the time and resources required for training a model from scratch, while enhancing precision. These models facilitate knowledge transfer across different domains, enabling more efficient and accurate solutions. The Grad-CAM algorithm serves as a visualization tool, elucidating the decision-making process of convolutional neural networks (CNNs). By utilizing gradients pertaining to a selected target class, Grad-CAM generates a localization map highlighting pivotal areas within an image critical for identifying a specific concept. This study introduces a comprehensive transfer model for synchronous COVID-19 detection and visualization of affected areas using X-ray imaging (He et al., 2016).

Unlike prior research, our work focuses on COVID-19 detection via X-ray images. Given the pressure on healthcare systems, it is crucial to leverage every available resource efficiently. Integrating the Grad-Cam localization feature aids in identifying COVID-19 and assessing severity, assisting in determining whether immediate healthcare intervention is needed (Selvaraju et al., 2016). In our investigation, we leveraged a diverse range of pre-trained models for the classification task, including:

The incorporation of Grad-CAM into our methodology enabled the precise identification of affected regions, with the EfficientNet architecture being utilized for its exceptional:

The utilization of pre-trained models offers significant advantages, primarily due to their pre-learned weights, which make them exceptionally valuable even when working with limited datasets. This constitutes the primary benefit of integrating pre-trained models into our investigative approach. Additionally, the employment of pre-trained models reduces the:

making them a computationally efficient option.

To address the variability in light intensity of the captured images, we applied the Contrast Limited Adaptive Histogram Equalization (CLAHE) technique to enhance image quality, ensuring a more accurate analysis. This preprocessing step enabled the:

ultimately leading to more reliable results.

In this study, we leveraged publicly accessible datasets of X-ray images as the primary experimental medium. The architecture of our experiment is designed as an end-to-end system, eliminating the need for manual feature extraction or selection, thereby streamlining the process for enhanced efficiency and effectiveness. We employed the Grad-CAM technique in conjunction with a Convolutional Neural Network (CNN) model to improve the diagnostic accuracy of our system. The integration of Grad-CAM enables the visualization of feature importance, allowing us to identify the most relevant regions in the X-ray images for diagnostic decision-making. By utilizing this approach, we aimed to develop a robust and accurate diagnostic system, capable of automatically detecting, and localizing abnormalities in X-ray images, thereby assisting clinicians in making informed decisions.

The collaborative efforts of Eduardo A. Soares, Plamen P. Angelov, and Sarah Biaso have culminated in a significant contribution to the field of medical imaging for infectious diseases. They have developed and made publicly available a comprehensive dataset of CT scans specific to SARS-CoV-2, enabling the advancement of diagnostic capabilities.

The team crafted an innovative algorithm, meticulously training it on CT scans from both:

This training phase was followed by a rigorous testing phase, where the algorithm's efficacy was validated on a distinct dataset encompassing CT scans from individuals with and without SARS-CoV-2 infection.

The outcomes of this testing phase demonstrated the algorithm's precision in accurately detecting SARS-CoV-2 infection, highlighting its potential utility in enhancing diagnostic processes within clinical settings. This contribution has the potential to significantly impact the field of medical imaging for infectious diseases, improving patient outcomes and streamlining diagnostic procedures (Soares et al., 2020). In a groundbreaking study, Sara Haseli and Nastaran Khalili have significantly advanced our understanding of COVID-19 pneumonia through comprehensive radiological analysis of chest CT imaging. Their research elucidated distinct hallmark features characteristic of the condition, including:

These findings align with established diagnostic criteria for COVID-19 pneumonia, providing critical insights into the disease's pulmonary manifestations.

Furthermore, their investigation revealed additional complications in a subset of patients, including:

These findings broaden our understanding of the disease's impact on pulmonary structures, underscoring the importance of vigilant radiological monitoring and timely diagnosis (Haseli et al., 2020). In terms of pulmonary involvement, the posterior segment of the left lower lobe (LLL) was identified as the most frequently affected segment, exhibiting a high propensity for involvement. Additionally, significant instances of involvement were observed in the right middle lobe (RML) and the right lower lobe (RLL), suggesting a bilateral distribution of pulmonary affliction. Upon analyzing the data based on lobar distribution, the LLL exhibited the highest frequency of affliction, with the RLL and RML closely following in prevalence, indicating a relatively even distribution of pulmonary involvement across the lobes (Li et al., 2020a).

Upon examining the demographics of age and gender in relation to chest CT imaging outcomes, a notable pattern emerged, suggesting a gender-specific predilection for lobar involvement. Male patients exhibited a significant propensity for left lower lobe (LLL) involvement, whereas female patients demonstrated a tendency toward more frequent involvement of the right lower lobe (RLL). Further analysis revealed a distinct age-related pattern, with the left lower lobe (LLL) being predominantly affected in the older population (65 years). In contrast, the right lower lobe (RLL) showed a higher incidence of involvement among younger individuals (<65 years). These findings suggest that age and gender may play a role in determining the lobar distribution of pulmonary involvement (Narin et al., 2021).

Wang et al. developed a cutting-edge algorithm for the detection of COVID-19 pneumonia via chest CT image analysis, leveraging an adapted Inception transfer-learning framework. The algorithm's performance was rigorously validated through both internal and external validation processes, demonstrating its efficacy in identifying COVID-19 pneumonia with a high degree of accuracy (Islam et al., 2020).

In their investigative study, Ali Narin and Ceren Kaya put forward three models grounded in convolutional neural network technologyaˆ C”ResNet50, InceptionV3, and InceptionResNetV2aˆ C”for the purpose of identifying coronavirus pneumonia from chest X-ray images. These models underwent rigorous evaluation on a dataset that included both confirmed COVID-19 cases and cases of conventional viral pneumonia, demonstrating the application of advanced deep learning techniques in the differentiation and diagnosis of respiratory illnesses (Saha et al., 2020; Wang et al., 2020; Song et al., 2021). In a pioneering approach, the developed system leveraged a hybrid architecture, synergistically integrating a Long Short-Term Memory (LSTM) classifier with a Convolutional Neural Network (CNN) dedicated to feature extraction and selection. The system's performance was rigorously evaluated using a dataset comprising 421 cases, including 141 instances with features indicative of COVID-19. Following the completion of its training phase, the model demonstrated exceptional performance capabilities, adeptly categorizing images as either COVID-19 positive or negative. The evaluation of the model's efficacy involved the application of 10-fold cross-validation, yielding an impressive accuracy rate of 97.3%. This remarkable achievement underscores the system's potential in medical imaging analysis, showcasing its ability to accurately classify images and support diagnostic decision-making (Mohammad and Abolfazl, 2020; Islam et al., 2020; Alharbi et al., 2022). Extensive validation was conducted on an independent dataset of chest X-ray images, where the proposed model achieved an exceptional accuracy of 97.7%. This remarkable performance demonstrates the model's precision in distinguishing COVID-19 cases, showcasing its potential in medical imaging analysis. The utilized dataset comprised 88 confirmed instances of COVID-19, 101 cases of bacterial pneumonia, and 86 instances classified as normal based on CT scan analyses. Comparative assessments were conducted to evaluate the model's performance relative to traditional frameworks, including Res-Net, Dense-Net, and VGG16. The results underscore the proposed model's enhanced performance, demonstrating its effectiveness through rigorous analysis (Singh et al., 2022; Alharbi et al., 2022a,b).

The allocation of the dataset for various purposes is segmented as follows:

Furthermore, within the Training + Validation subset, the images are divided into:

The distribution and representation of all three subsets (COVID-19, Normal, and Pneumonia) are illustrated in Figures 2, 3.

The synergistic integration of Convolutional Vision Transformers (CVT) and Convolutional Channel Transformers (CCT) represents a groundbreaking approach in object recognition, harnessing the complementary strengths of Convolutional Neural Networks (CNNs) and transformers to process images with enhanced efficacy. This innovative methodology enables a comprehensive and holistic analysis of images, significantly improving the model's efficiency and deployment capability. The remarkable performance of this approach is particularly evident in the classification of COVID-19 images, as illustrated in Figure 8 respectively. By combining the spatial hierarchies of CNNs with the self-attention mechanisms of transformers, CVT and CCT facilitate a more detailed and nuanced understanding of image features, leading to improved recognition accuracy and robustness. This integrated approach demonstrates a significant advancement in computer vision, enabling more effective and efficient image analysis in various applications, including medical imaging and disease.

The metrics are defined as follows, where TP represents true positives, TN denotes true negatives, FP stands for false positives, and FN signifies false negatives:

The primary objective of this experiment was to evaluate the performance of classification models on a dataset of X-ray images, encompassing categories such as standard, lung opacity, pneumonia, and COVID-19 cases. To ensure optimal model training and evaluation, data enhancement and balancing techniques were applied, as illustrated in Tables 1, 2.

To ensure fair and reliable model predictions, techniques were applied to address class imbalance in the dataset, which could otherwise bias the model toward overrepresented classes. Methods such as class weighting and oversampling were employed to balance the distribution among the COVID-19, pneumonia, and normal classes. Class weighting adjusted the loss function to give higher importance to minority classes, while oversampling involved duplicating samples from underrepresented classes to create a more balanced dataset. These approaches aimed to reduce bias and improve the model's ability to accurately classify images across all categories, enhancing its reliability and robustness in clinical applications.

The initial phase of our image data pre-processing involved two critical steps: image enhancement using CLAHE and subsequent data augmentation. This comprehensive approach was designed to improve the quality and variability of the dataset, thereby aiding in the robustness of the subsequent classification models. To boost the quality of input images and enhance model accuracy, a technique called Contrast Limited Adaptive Histogram Equalization (CLAHE) was used to refine each image. This step improves contrast and highlights important features within X-ray images, leading to more accurate predictions. Additionally, image resolution standardization was performed to ensure consistent image sizes, making the model adaptable to various image sources and minimizing potential variability from different imaging devices. These preprocessing steps lead to a more robust model that can generalize across diverse imaging conditions, ensuring reliable performance in real-world applications.

The datasets were subjected to preprocessing using Contrast Limited Adaptive Histogram Equalization (CLAHE), followed by division into two distinct sets: the original set and the balanced set, with and without additional enhancement. A comprehensive evaluation of various models was conducted on these datasets, assessing their performance based on accuracy and loss metrics across three phases:

This rigorous evaluation framework enabled a thorough analysis of model performance, facilitating the identification of optimal models and hyper parameters for the task at hand.

To prevent overfitting and enhance the model's generalizability, several techniques were employed during training. First, cross-validation was used to ensure robust model evaluation across different data splits, which helped identify any potential overfitting to specific subsets. Additionally, data augmentation techniques such as random rotations, translations, and flips were applied to increase dataset variability and reduce the model's reliance on specific image features. Dropout layers were also incorporated within the model architecture to prevent neurons from co-adapting too strongly, which often leads to overfitting. Despite these measures, validating the model on external datasets is crucial to further assess its adaptability and effectiveness across varied real-world settings and populations. This will be a key focus of future work, as it is essential to ensure that the model can generalize well to new, unseen data, and provide accurate predictions for a diverse range of patients and scenarios.

The models evaluated included Xception, InceptionV3, and InceptionResNetV2. The performance metrics revealed variations in accuracy and loss across the datasets (Table 3).

Disease localization in medical images is a critical step in diagnostic processes. The application of Grad-CAM to models trained on various datasets elucidates the relationship between training data quality, model accuracy, image resolution, and localization precision. Grad-CAM uses the gradients of any target concept, flowing into the final convolutional layer to produce a coarse localization map highlighting the important regions for predicting the concept. Mathematically, it can be represented as

Where LGrad-CAMC is is the localization map for class c, αkc are the weights for feature map k, A^k is the activation of k-th feature map, ReLU is applied to focus on features that have positive influence on the class of interest. Results the application of Grad-CAM on models trained with initial and enhanced 478 datasets revealed that models trained on initial data more accurately highlighted disease-affected areas. This accuracy in localization is directly proportional to the model's overall accuracy and the image resolution, described as:

emphasizing the compounded effect of higher accuracy and better resolution on precise disease localization. Our findings underscore the importance of image quality and model accuracy for effective disease localization using Grad-CAM. The study advocates for the optimization of these factors to improve diagnostic efficiency in medical imaging as shown in Figures 12, 13.

While Grad-CAM provides valuable insights into the decision-making process behind COVID-19 predictions, deep learning models are often criticized for their lack of transparency. This “black box” nature can be a significant barrier for clinical adoption, as clinicians may struggle to understand the reasoning behind model predictions. To address this, future work could explore combining Grad-CAM with other explainability techniques, such as LIME or SHAP. These methods offer unique perspectives on model behavior, providing clinicians with a more comprehensive understanding of prediction rationales. By shedding light on the decision-making process, we can increase trust and usability in medical settings.

The advent of deep learning in medical imaging has facilitated the development of automated diagnostic tools. This paper presents an evaluation of transfer learning models, specifically EfficientNet and MobileNet, in the classification of chest X-ray images. Transfer learning models were trained on a comprehensive dataset comprising images categorized as COVID-19, normal, and viral pneumonia. The performance was assessed based on the accuracy of classifications, with further analysis conducted through confusion matrices and ROC curves.

In response to the COVID-19 pandemic, our study demonstrates promising results for COVID-19 detection using chest X-rays. However, we must note that our model has not yet been tested in real-world clinical settings, which limits our ability to fully assess its performance in a practical healthcare environment. To address this, we plan to validate our model in clinical settings to evaluate its effectiveness, robustness, and potential impact on patient diagnosis and care. This will provide valuable insights into our model's applicability in different medical scenarios and move us closer to broader adoption in clinical practice. Our proposed solution leverages the power of pre-trained models and demonstrates commendable efficacy, achieving an accuracy rate of 88.48% in training and 88.1% in validation on the initial dataset. By harnessing Efficient Net-based transfer learning on a balanced and enhanced dataset, our developed models have attained exemplary performance, registering training and validation accuracies of 95.64% and 97.31%, respectively. These results not only parallel but, in some instances, surpass the accuracy levels of existing models, demonstrating the robustness of our approach. Notably, our models' enhanced capability to precisely localize affected areas significantly bolsters their diagnostic utility, providing a valuable tool for physicians in the fight against COVID-19. Our study contributes to the growing body of research in AI-assisted medical imaging, showcasing the potential of deep learning architectures to revolutionize healthcare diagnostics.

The future directions section would benefit from a more comprehensive roadmap. Specifically, the paper should address several promising avenues: (1) the integration of advanced explainability techniques like SHAP or LIME to enhance model interpretability; (2) validation across diverse datasets from different domains to establish broader generalizability; and (3) exploration of hybrid approaches combining the current method with emerging techniques in the field. These extensions could address current limitations while advancing the broader research agenda.

First, while our dataset includes chest X-ray images, it may not fully represent the diversity of COVID-19 cases across different populations and imaging equipment. The model, though showing high accuracy in experimental settings, requires validation in real-world clinical environments to establish practical utility. While we employed Grad-CAM for visualization, we recognize that our model's interpretability could be enhanced through additional techniques like SHAP or LIME to increase clinician trust. Despite implementing class weighting and oversampling, inherent dataset imbalances persist, potentially affecting prediction reliability for minority classes. The model's computational requirements may pose challenges in resource-constrained settings, suggesting a need for architectural optimization. Although we implemented dropout and data augmentation, the high-test accuracy warrants external validation to conclusively demonstrate generalizability. Finally, our focus on X-ray imaging alone may not capture all relevant COVID-19 clinical features, indicating potential value in incorporating additional imaging modalities like CT scans or clinical data in future work.