The rapid expansion of telemedicine platforms enabled remote assessment and monitoring of COVID-19 patients, especially during peak transmission periods when hospital resources were overwhelmed [13]. However, telehealth often lacks the diagnostic depth provided by imaging or laboratory testing and is generally used for symptom tracking and triage rather than precise diagnosis.

Chest X-rays and CT scans have been instrumental in identifying characteristic COVID-19 lung involvement, including bilateral ground-glass opacities and consolidations [14]. Numerous deep learning models have been developed for pneumonia and COVID-19 detection using chest X-ray and CT data. For example, MobileNet (mobile network) achieved 94.2% and 93.7% accuracy on 2 public chest X-ray datasets containing 5856 and 112,120 images, respectively [15]. Despite these benefits, existing studies often suffer from limited and nonstandardized datasets, a lack of demographic metadata (age and sex), and geographical imbalance, reducing generalizability. In a separate study using InceptionV3 and convolutional neural network (CNN) models on a Kaggle X-ray dataset of 7750 images, the researchers reported impressive results (accuracy: 99.2%, recall: 99.7%) [16]. However, the use of a single public dataset lacking demographic diversity and external validation limits generalizability.

A CT-based study using NASNet achieved an exceptionally high accuracy of 99.6%, with a sensitivity of 99.9% and a specificity of 98.6% [17]. However, this evaluation was based on a small, imbalanced dataset of 249 patients, with no external validation, no interpretability tools, and no metadata analysis (eg, age, sex, and geography), weakening its clinical reliability and fairness. Furthermore, alternative architectures like ResNet or VGG were not benchmarked, and hyperparameter tuning was minimally discussed.

These limitations underscore the need for scalable, diverse, and metadata-rich imaging datasets to enhance model reliability and cross-population performance.

While reverse transcription–quantitative polymerase chain reaction (RT-PCR) remains the diagnostic gold standard [18], its accuracy can be impacted by emerging variants and sample quality. In response, several alternative diagnostic technologies have been explored. A comparison of key methods is presented in Table 3.

PCR-based methods are highly accurate but not variant-agnostic. Antigen-based tests are accessible but less reliable. Genome sequencing is ideal for surveillance but not rapid diagnosis. These constraints further support the need for AI-powered imaging diagnostics that are scalable, noninvasive, and rapid.

Deep learning applied to medical imaging presents a promising complementary diagnostic method, particularly in areas with limited laboratory capacity. Yet, current research has notable limitations. For instance, a protocol paper of a prospective AI model for chest X-ray images highlights the intention to use 600 images [19]. However, it lacks clear details on geographic and demographic diversity, metadata tracking (eg, age and sex), and model architecture. Moreover, it does not describe how biases will be addressed or how low-prevalence conditions will be handled, which can be considered critical for real-world implementation.

Given the diagnostic delays and limitations associated with conventional methods, deep learning applied to medical imaging offers a promising complementary approach. Models trained on chest X-rays and CT scans can provide rapid, accurate, and interpretable results, which are particularly critical in settings where molecular testing is delayed or inaccessible. In this study, these efforts were built upon by employing transfer learning on an expanded, standardized imaging dataset to enhance diagnostic accuracy and generalizability. This approach addresses prior limitations related to data volume, diversity, and model robustness.

New SARS-CoV-2 variants, such as Pi, Rho, XEC, and JN.1, exhibit mutations in the spike (S) and nucleocapsid (N) proteins, which impair molecular and antigen-based diagnostic assays [20]. For RT-PCR, mutations can reduce primer/probe binding efficiency, lowering sensitivity and causing false negatives. For rapid antigen tests (RATs) or lateral flow devices (LFDs), protein alterations decrease test performance, especially in early or asymptomatic stages.

Radiological signs of COVID-19 (ground-glass opacities) overlap with other pulmonary infections, including bacterial pneumonia, influenza, tuberculosis, respiratory syncytial virus, and fungal infections. This nonspecificity complicates diagnosis, especially without clinical or laboratory correlation, increasing the risk of false positives or misclassification.

Many existing AI models are trained on limited or biased datasets, which can impact their generalizability. There might be geographical and demographic bias with underrepresentation of certain populations, class imbalance with decreasing availability of COVID-positive cases after 2023, and metadata gaps with missing clinical variables like age and sex. These limitations reduce model robustness, especially in real-world settings with varied patient populations.

Despite promising research, AI tools face challenges in clinical adoption, including a lack of regulatory validation (Food and Drug Administration approval/Conformité Européenne certification), poor integration with electronic health records (EHRs), and clinician skepticism due to a lack of explainability or interpretability. Without improved trust, transparency, and workflow compatibility, real-world deployment remains limited.

Privacy regulations (Health Insurance Portability and Accountability Act and General Data Protection Regulation) and institutional data silos restrict access to multicenter, diverse datasets and large-scale, cross-border collaborations necessary for robust AI development.

Most diagnostic tools are optimized for acute-phase detection. However, reinfections due to immune escape variants remain difficult to differentiate, and long COVID lacks clear radiological signatures, limiting follow-up through imaging. There is a need for diagnostic systems that can also support longitudinal patient monitoring.

Low-income regions often lack access to RT-PCR labs, CT or X-ray imaging facilities, and high-performance computing resources for AI deployment. This exacerbates health inequities and delays early detection and containment efforts.

Unlike RT-PCR and antigen tests that rely on viral RNA or surface protein stability, the present image-based approach detects disease-induced radiological changes, remaining unaffected by emerging variants or antigenic drift.

Imaging-based models do not depend on spike or nucleocapsid protein integrity, making them robust against variants like XEC and JN.1.

Transfer learning has been adopted using pretrained CNNs on ImageNet, and they have been fine-tuned on curated COVID-19 datasets with advanced preprocessing, augmentation, and optimization strategies.

The system is designed for binary classification (COVID-19 vs normal) and multiclass classification (COVID-19 pneumonia vs non-COVID pneumonia vs normal), depending on available label granularity. Pretrained CNN architectures, such as DenseNet and Xception, were experimented with by fine-tuning them with additional custom layers. The models were further optimized through hyperparameter tuning, and attention modules were incorporated to improve the network’s ability to focus on COVID-relevant regions in the lung fields.

To improve generalization, a dataset of 25,195 labeled images has been assembled across CT and X-ray modalities; multiple regions (Asia, Europe, and North America); and varying age groups, ethnicities, and imaging protocols. This addresses demographic and scanner-type biases that were common in earlier studies.

Grad-CAM visualizations have been integrated for transparent decision support.

The present framework has been designed to be extended for follow-up analysis, allowing radiological tracking of postinfection abnormalities and aiding in long COVID assessment and reinfection detection.

The final model has been compressed using quantization and pruning techniques for deployment in edge devices (mobile apps and local hospital servers) and cloud-assisted diagnostic platforms.

Traditional approaches like RT-PCR, LFDs, and RATs, though widely used, suffer from several drawbacks: reduced sensitivity with emerging variants due to mutations in target genes and proteins; delayed turnaround times in lab-based settings; sample quality dependency leading to false negatives, especially in asymptomatic individuals; and lower reliability in detecting newer variants such as Pi, Rho, XEC, and JN.1. These limitations necessitate complementary, mutation-resilient diagnostic strategies.

Chest CT scans and X-rays have proven valuable in identifying COVID-19–induced pneumonia, with CT offering higher sensitivity (88%‐97%) and X-rays being cost-effective and more widely available, especially in resource-constrained environments [4].

The application of deep learning and transfer learning to radiological image analysis enhances diagnostic accuracy, speed, and consistency, independent of viral genome variability or test kit supply chains.

This study developed and evaluated a deep learning diagnostic framework using CT and X-ray images to detect COVID-19 pneumonia. The key goals were to achieve a diagnostic accuracy of >95% across multiple viral variants; improve generalization across populations, regions, and imaging devices; differentiate COVID-19 pneumonia from other respiratory conditions with overlapping features; and benchmark the model’s performance against traditional diagnostic methods.

To ensure clinical reliability, the model must distinguish COVID-19 pneumonia from visually similar conditions. The radiological overlap emphasizes the need for fine-grained classification models capable of accurately distinguishing COVID-19 from similar pulmonary pathologies using feature-rich image interpretation.

This study aimed to develop a mutation-resilient deep learning framework for accurate COVID-19 diagnosis using CT and X-ray imaging, overcoming challenges faced by traditional RT-PCR and antigen tests due to emerging SARS-CoV-2 variants. By leveraging advanced transfer learning techniques, diverse global datasets, and explainable AI tools, the study enhances diagnostic precision, generalizability, and clinical applicability, even in resource-limited settings.

The dataset comprised radiological data from 9 primary sources. Each source was selected based on the following inclusion criteria: confirmed diagnostic status, with only RT-PCR–confirmed COVID-19 cases and clinically validated normal or pneumonia samples included; radiological quality, with DICOM or high-resolution image formats (PNG and JPEG) and clear lung visibility; and metadata completeness, with availability of patient demographics (age and sex), scan modality, and clinical context, where applicable.

The following imaging datasets were considered:

To correct for imbalance and ensure representative learning, undersampling of Spain was performed to reduce overrepresentation, and countries with fewer than 100 total samples were removed to prevent noise and overfitting.

Significant missing values were found in the metadata. Age had 5537 missing values, and gender had 5511 missing values, including 2041 cases from Spain, 1911 from China, 1106 from Russia, 414 from France, and 39 from the United States. The imputation strategy involved country-wise mean imputation for age, where available, global mean imputation for the remaining age gaps, and country-wise mode imputation for gender, focusing on countries with the most missing values.

The age range was 0 to 100 years. Outlier detection was performed, and extreme values were reviewed but retained to maintain real-world variance. Patients were categorized into discrete age groups (eg, 0‐18, 19‐35, 36‐60, and 61+ years), allowing demographic stratification during training. To handle age group imbalance during dataset splitting, the stratify label by age group was applied.

The number of final images after metadata curation was 11,052 (8842 for training and 2210 for validation). The preprocessing pipeline included image resizing to 75×75 pixels with 3 channels (RGB) and normalization with pixel values rescaled to [0, 1].

The distribution of COVID-19 and normal images is presented in Multimedia Appendix 1. Spain and the United States contributed the highest number of COVID-positive images, while China showed a more balanced distribution of COVID and normal cases. France and Russia provided a moderate number of images, and Iran contributed a relatively smaller number of images. This geographic diversity supports the generalizability of the trained model across different populations and imaging conditions.

To balance samples across underrepresented countries, the following augmentation techniques were applied: random horizontal flip, random rotation (15°), random zoom (10%), random contrast (10%), and random translation (5%).

Despite country-level augmentation, class imbalance between the COVID-19 and normal categories persisted. Additional category-level augmentation was applied to underrepresented normal samples to achieve closer class parity, helping reduce bias during model training.

After applying data augmentation techniques, the final dataset consisted of 24,408 medical images, which were stratified to maintain balanced class distributions across all subsets. The dataset was divided into 19,527 images for training, 4881 for validation, and 952 for testing. Stratified sampling ensured proportional representation of each class, supporting fair evaluation and reducing potential bias during model training and validation.

All images were resized to 224×224 pixels to ensure consistent input dimensions compatible with standard CNNs. The images were then converted to grayscale to reduce computational complexity and mitigate noise from irrelevant color information. Pixel intensities were normalized to stabilize training dynamics.

To determine an optimal batch size for training, an analysis was performed regarding how different batch sizes divide the total training dataset of 19,527 records. This involved calculating how many steps (batches) each epoch would require for various batch sizes. Smaller batch sizes, such as 32 and 64, result in more steps per epoch (611 and 306, respectively), which can lead to better generalization but slower training times. On the other hand, very large batch sizes like 512 or 1024 reduce the number of steps significantly but may hinder model generalization and require careful tuning of the learning rate. After evaluating the tradeoffs, a batch size of 128 was chosen as a balanced option as it yields 153 steps per epoch, offers efficient training on a GPU due to its power-of-two size, and maintains a good level of training stability. This choice reflects a compromise between computational efficiency and model performance, ensuring the training process remains both practical and effective.

To address class imbalance, a combination of data augmentation and undersampling strategies was implemented. The dataset was split into 80% for training and 20% for validation, and performance was further optimized using caching and shuffling for the training set. For the validation set, caching alone was applied to ensure consistent evaluation.

To enhance the randomness of the training data, the buffer size was set to 10,000 during the shuffling process. The buffer size determines how many samples are held in memory and randomly shuffled at any given time before being passed to the model in batches. A smaller buffer size, such as 100 or 1000, can result in less effective shuffling, especially with larger datasets, as only a limited portion of the data is randomly sampled at a time. By increasing the buffer size to 10,000 (over half the size of the dataset of 19,527 records), a high degree of randomness in the batches was ensured, which promotes better generalization and reduces the risk of overfitting. Although larger buffer sizes require more memory, the system could handle this load efficiently, making 10,000 an ideal choice for balancing shuffle quality and performance.

A structured and modular deep learning pipeline was developed for hyperparameter optimization and fine-tuning using TensorFlow and Keras Tuner. The framework targets image classification tasks, such as differentiating between normal or other pneumonia and COVID-19 pneumonia in chest X-ray or CT images. The pipeline combines automated hyperparameter tuning, transfer learning, and robust training strategies to improve classification accuracy and generalization, which are particularly crucial when dealing with limited medical datasets.

The model was trained over 30 epochs with a batch size of 128, a buffer size of 10,000, and a fixed random seed of 42 to ensure reproducibility.

To determine the optimal number of training epochs without overfitting, early stopping was used, which is a regularization technique that monitors validation performance during training. Instead of predefining a fixed number of epochs, early stopping halts training once the validation loss stops improving for a set number of consecutive epochs (patience). This dynamic approach allows the model to train just long enough to reach optimal performance without wasting computation or risking overfitting. Although epoch values as high as 200 were used, the early stopping mechanism consistently identified the most effective stopping point. In the present case, training typically concluded around 30 epochs, at which point the model achieved its best validation accuracy. This method provided an efficient and reliable way to control training duration while ensuring strong generalization.

At the core of the architecture was a transfer learning model based on VGG16 (Visual Geometry Group network-16 layers), which was selected as the baseline due to its simple, deep CNN structure consisting of 16 layers with repeatable 3×3 convolution and max-pooling blocks. VGG16 is well-established in medical imaging research and serves as a strong, interpretable starting point.

To determine the most effective transfer learning strategy, various freeze rates of 0.01, 0.05, 0.10, 0.20, 0.50, and 0.75 were considered, and the following formula was used to calculate how many layers of the pretrained base model to freeze: num_freeze_layer = int(len(base_model.layers) × freeze_rate).

The freeze rate controls how much of the original model’s learned features are retained versus fine-tuned on the new task. In general, higher freeze rates, such as 0.50 or 0.75, are preferable when working with small datasets or datasets like the original training data (ImageNet), as they help prevent overfitting and preserve general visual features. Conversely, lower freeze rates, such as 0.01 or 0.05, are more suitable for large or highly domain-specific datasets, where extensive fine-tuning is necessary. For many practical applications, mid-range freeze rates like 0.10 or 0.20 often provide the best balance, allowing the model to adapt to new data while still leveraging pretrained knowledge effectively.

Most layers of the pretrained model were frozen, except for selected unfrozen layers, enabling selective fine-tuning to adapt high-level features to the target domain while preserving learned representations.

As part of the model architecture, a GlobalAveragePooling2D layer was incorporated after the convolutional base. This layer plays a crucial role in reducing the spatial dimensions of the feature maps while preserving the most important information. Unlike traditional flattening, which converts the entire feature map into a long vector (often leading to many parameters), GlobalAveragePooling2D computes the average of each feature map, resulting in a much more compact representation. This not only reduces the risk of overfitting but also maintains the model’s spatial awareness and generalization ability. Additionally, it helps bridge the convolutional layers and the dense output layer in a more efficient and scalable way, especially when working with transfer learning models.

To further mitigate overfitting and improve generalization, a Dropout layer was added after the GlobalAveragePooling2D layer. Dropout works by randomly setting a fraction of the input units to zero during training, which prevents the model from becoming too reliant on specific neurons. Several dropout rates (0.2, 0.3, 0.4, and 0.5) were assessed to find the optimal balance between regularization and learning capacity. Lower dropout rates like 0.2 provided lighter regularization and allowed the model to retain more features, while higher rates like 0.5 offered stronger regularization but at the cost of slower learning. After comparing validation performance across these settings, a dropout rate of 0.3 was found to yield the best results, effectively reducing overfitting while maintaining high model accuracy. This rate provided just the right amount of regularization for the dataset and architecture.

Although the input dataset was prenormalized, a BatchNormalization layer was still incorporated within the model architecture. While input normalization standardizes the data fed into the model, BatchNormalization operates between layers, dynamically normalizing the activations during training. This helps address internal covariate shift, where the distribution of layer inputs changes due to updates in earlier layers, thus stabilizing training, enabling higher learning rates, and often improving generalization. Even with normalized input data, this internal normalization contributed to faster convergence and improved validation performance across experiments.

To determine the ideal size for the fully connected (dense) layer, various unit sizes (32, 64, 128, 256, and 512) were assessed. The number of units in the dense layer directly impacts the model’s ability to learn complex patterns. Smaller sizes like 32 or 64 limit the model’s capacity and are often suitable for simpler tasks or small datasets. Larger sizes like 256 or 512 increase representational power but also introduce a greater risk of overfitting, especially if the dataset is not sufficiently large or diverse. It was observed that as the number of units increased, the model’s ability to capture nuanced patterns improved up to a point. Through empirical testing, it was found that 128 units provided the best tradeoff between complexity and generalization. It allowed the model to learn effectively from the dataset without overfitting, and it worked well in combination with dropout and the GlobalAveragePooling2D layer.

To assess the real-time applicability of our target system, 2 model architectures were compared to balance performance and efficiency. Both began with a pretrained base model, followed by GlobalAveragePooling2D, BatchNormalization, and an initial Dropout and Dense layer. The first architecture included an additional Dropout and Dense layer, designed to improve representational capacity and regularization. The second architecture was more streamlined, using only a single Dropout and Dense layer before the output.

In the context of real-time deployment, model efficiency is crucial. While the deeper architecture offered slightly better training performance, it came at the cost of increased latency and model complexity. Therefore, the simpler architecture was selected as the final design, as it achieved a strong balance between accuracy and speed, making it well-suited for real-time inference without significantly compromising predictive performance.

The Dense layer had rectified linear unit activation, He-normal initialization, and L2 regularization. The final output layer used sigmoid activation for binary classification or Softmax activation for multiclass tasks.

As part of the optimization strategy, several well-known optimizers, including SGD, RMSprop, Adam, Nadam, and AdamW, were evaluated. Each optimizer has unique strengths: SGD offers strong theoretical foundations but typically requires fine-tuned hyperparameters; RMSprop is effective in handling nonstationary objectives; Adam combines momentum and adaptive learning rates, leading to fast convergence; and Nadam incorporates Nesterov momentum into Adam for smoother updates. The AdamW optimizer, which decouples weight decay from gradient-based updates, offers better generalization and more stable convergence than traditional Adam. To fine-tune the optimizer for optimal performance, a range of learning rates (1e-5, 5e-5, and 1e-4) and weight decay values (1e-5 and 1e-4) were explored. This tuning allowed the model to adapt effectively to the complexity of the dataset while minimizing overfitting. After extensive experimentation, it was found that a learning rate of 5e-5 combined with a weight decay of 1e-5 yielded the best results, providing smooth convergence, strong validation accuracy, and robust generalization. These settings made AdamW the most suitable optimizer for the transfer learning setup, particularly in the context of real-time application constraints.

Binary cross-entropy was used as the loss function for binary classification, while categorical cross-entropy was employed for multiclass settings. Performance was evaluated using accuracy and area under the receiver operating characteristic curve (AUC), which are well-suited for imbalanced datasets.

To enhance training efficiency and prevent overfitting, several callbacks were incorporated. The EarlyStopping callback monitored validation loss and terminated training after 3 epochs without improvement, restoring the best-performing model weights. ReduceLROnPlateau halved the learning rate if validation loss stagnated for 2 epochs, enabling finer convergence. A model checkpointing strategy saved the full model, including weights and architecture, to a specified directory at each epoch, regardless of validation performance, ensuring training continuity and recovery if interrupted.

Automated hyperparameter optimization was performed using the Hyperband algorithm implemented in Keras Tuner. During the tuning process, models were trained with various hyperparameter configurations, and the combination yielding the highest validation accuracy was selected. Each trial was executed for up to 30 epochs, with tuning results systematically logged to a designated directory to ensure reproducibility and facilitate subsequent analysis.

The tuning process was orchestrated by a centralized function that built the model based on sampled hyperparameters, applied callbacks, conducted training on the training and validation splits, and identified the best-performing configuration. The final model, constructed using this optimal configuration, was retrained on the full training data and saved for future deployment or evaluation.

This framework offers several key advantages. It automates the search for critical hyperparameters, such as dropout rates, dense layer sizes, and learning rates, reducing the reliance on manual tuning. Leveraging pretrained models improves learning efficiency and generalization, which is particularly valuable when working with small or noisy medical datasets. Furthermore, the integration of early stopping, adaptive learning rate scheduling, and model checkpointing ensures robust, reliable training. Collectively, these strategies contribute to the development of accurate and generalizable deep learning models suitable for real-world clinical applications.

The formula for accuracy is as follows: accuracy = (true positive + true negative) / (true positive + true negative + false positive + false negative). Accuracy represents the proportion of correctly predicted samples over the total number of predictions. It is a suitable metric when the dataset is balanced across classes.

The formula for precision is as follows: precision = (true positive) / (true positive + false positive). Precision measures the correctness of positive predictions. It is especially important when the cost of false positives is high.

The formula for recall is as follows: recall = (true positive) / (true positive + false negative). Recall assesses the model’s ability to identify actual positives. It is critical in scenarios like medical diagnosis, where missing positive cases can have serious consequences.

The formula for F₁-score is as follows: F₁-score = 2 × ([precision × recall] / [precision + recall]). The F₁-score balances precision and recall and is particularly useful when working with imbalanced datasets.

The receiver operating characteristic curve plots the true positive rate (recall) against the false positive rate. The AUC represents the probability that a randomly chosen positive instance is ranked higher than a randomly chosen negative instance. A higher AUC indicates better model discrimination capability.

The curated dataset, representing emerging variants, such as Pi, Rho, Xec, and JN.1, enabled model training and validation with high precision.

The deep learning model outperformed traditional tests in sensitivity for variant cases. For instance, while RT-PCR sensitivity dropped for Pi and JN.1, the model maintained >98% recall in cross-validation trials.

By incorporating images from 19 countries across different imaging modalities and population groups, the model exhibited stable performance across validation subsets with different geographic and demographic characteristics.

Fine-grained classification enabled the model to distinguish COVID-19 pneumonia from other respiratory infections (bacterial and atypical pneumonias), achieving a specificity of 96.9% and an F₁-score of 97.9%.

A diverse set of imaging datasets spanning CT and X-ray modalities was compiled from multiple countries to ensure model generalizability and robustness. A total of 87,231 images were identified. The largest contributor was BIMCV-COVID19 from Spain with 79,023 (90.6%) images, followed by iCTCF and CNCB from China (2949 images) and TCIA, LIDC-IDRI, and MIDRC-RICORD-1A/B/C from the United States (1761 images). Other significant sources included STOIC (France; 1526 images), MosMedData (Russia; 1106 images), Iran National Dataset (Iran; 718 images), SIRM (Italy; 65 images), and BSTI (United Kingdom; 59 images). Additionally, radiological images were extracted from global resources like Radiopaedia and contributions from 11 other countries, each providing 24 cases (Multimedia Appendix 2). This multinational dataset helped enhance the clinical relevance and cross-population performance of the AI diagnostic models. BIMCV-COVID19 (Spain) contributed the largest number of both positive and negative samples, and there were smaller contributions from datasets such as SIRM (Italy), CHQC (China), and MIDRC-RICORD (United States) (Multimedia Appendix 3). The distribution highlights the dataset’s diversity and the class balance achieved across sources, which are critical for training robust and unbiased diagnostic models.

Most data were collected from the BIMCV-COVID19 dataset (Spain), which, while enhancing the dataset’s size and regional representation, introduces a notable class imbalance. Specifically, COVID-19 positive cases (59,961) substantially outnumbered normal and non-COVID pneumonia cases (27,270). This disproportion primarily stems from the emphasis of public datasets on rapid COVID-specific data collection during the pandemic, which may skew model learning and diagnostic performance if not addressed.

After applying undersampling and dropping countries with ultra-low samples, 3000 cases from Spain, 2949 from China, 1761 from the United States, 1526 from France, 1106 from Russia, and 718 from Iran were retained, with 7572 COVID cases and 3488 non-COVID cases.

Age had 5537 missing values imputed using country-wise medians, and gender had 5511 missing values imputed using country-wise modes. Metadata completeness improved to 100%, allowing demographic-aware stratification and analysis during model evaluation.

The age range was 0 to 100 years. After removing outliers, the age group distribution remained imbalanced, with adults (n=7219) forming the majority, followed by elderly (n=2234), young adults (n=1553), and children (n=54). The distribution reflects a population skew that may influence age-specific modeling outcomes, and stratified labels by age group ensure balanced data. Gender balance included 34.1% (3398/9954) males and 65.9% (6556/9954) females. After processing, the dataset had 4509 positive and 2047 negative cases among females, and 2307 positive and 1091 negative cases among males. This balanced demographic composition supports robust model evaluation across diverse patient profiles.

After applying country-based filtering, undersampling, and augmentation, a more equitable distribution of samples across countries and classes was achieved. The total number of curated images was 11,052, with 8842 images in the training set and 2210 images in the validation set. Image dimensions were resized to 75×75 pixels with 3 RGB channels, and normalization was applied with all pixel values rescaled to the [0, 1] range.

Augmentation techniques were applied particularly to underrepresented classes to reduce class imbalance and enhance model generalization. A balanced representation (2034 COVID samples per country) was achieved across 6 key contributors (China, France, Iran, Russia, Spain, and the United States). Similarly, normal samples were balanced at 1249 images across the same regions, improving generalization across populations (Multimedia Appendix 4).

The dataset comprised 12,204 COVID-19–positive images and 7494 normal images, indicating a moderate class imbalance favoring positive cases. This distribution highlights the need for balancing techniques such as augmentation during model training.

The applied augmentation techniques (flip, rotate, zoom, contrast, and translation) not only balanced the dataset but also increased image variability, simulating real-world noise and improving model resilience to unseen data (Multimedia Appendix 5). There was a nearly equal number of images per label (nearly 2000 per class) in each country, demonstrating successful class balancing to mitigate bias during model training.

There was an equal number of COVID-positive and normal (COVID-negative) images (12,204 each), reflecting the successful application of augmentation techniques to balance the dataset and prevent model bias due to class imbalance. Class distribution after augmentation is presented in Multimedia Appendix 6.

There is an aim to collaborate with multiple hospitals and diagnostic centers to externally validate the model on institution-specific datasets. This will help assess the model’s generalizability and robustness across different scanners, protocols, and patient populations.

Work is underway to integrate the diagnostic tool with EHR systems for seamless access to patient history and real-time imaging data, enabling context-aware predictions and decision support.

The final model is being optimized using techniques, such as quantization and pruning, for deployment on edge devices and cloud platforms. This will support real-time diagnosis via a web interface and mobile app, particularly in resource-constrained or rural areas.

Documentation and performance benchmarks are being prepared to pursue regulatory approval (Conformité Européenne marking and Food and Drug Administration clearance). A prospective clinical trial is also being designed to measure diagnostic impact in a real-world setting.

There is a plan to adapt the system for longitudinal analysis, enabling clinicians to track radiological changes over time, which can be useful for monitoring long COVID progression or reinfections.

To support data privacy and multi-institutional collaboration, an attempt will be made to explore federated learning frameworks that allow model training on decentralized data without sharing patient images.