Cutting-edge bayesian deep learning and statistical strategies for bias mitigation in COVID-19 detection via chest x-ray imaging

Abstract

Chest radiography (CXR is widely used for triage and follow-up of pulmonary disease, yet COVID-19 classification remains vulnerable to bias, label noise, and domain shift. We propose a multi-stage Bayesian deep learning framework that combines lung segmentation, segmentation-guided classification, calibrated ensembling, and uncertainty estimation to classify four classes (COVID-19, normal, viral pneumonia, bacterial pneumonia) and to grade COVID-19 severity. Models are trained and tested on 1,531 CXRs (100 COVID-19 images from 70 patients; 1,431 non-COVID images from ChestX-ray14) with patient-wise splits. The final ensemble achieves 98.33% test accuracy; COVID-19 sensitivity reaches 100% on this split. Robustness is quantified by stress-testing five image degradations (Gaussian noise, motion/defocus blur, JPEG compression, and downsampling), with macro AUC drops remaining small at moderate severities and larger under strong blur or heavy downsampling. Saliency and context-relevance analyses are used to identify spurious cues. The study is limited by dataset size and lack of external multi-site validation; a planned evaluation on COVIDx and BIMCV-COVID19 + is outlined.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-28723-x.

Introduction

The COVID-19 pandemic, a relentless and untreatable disease, persists in causing fatalities worldwide, demanding immediate and critical attention. Although viruses such as the flu usually appear during a period of ten years, their existence is brief in comparison to the long-lasting presence of COVID-19. Despite extensive endeavors, only a small number of vaccines have been created as a result of restricted accessibility. The widespread impact of the pandemic and the lack of a universally effective vaccination emphasize the immediate necessity for answers. Emerging potential treatments, such as plasma therapy and X-ray imaging, have been identified, but a clear cure for the condition is yet unknown^1,2. The profound repercussions on individuals, economies, and healthcare systems emphasize the pressing need to discover a solution.

Many CXR classification studies were trained and evaluated on small, heterogeneous datasets with image-level rather than patient-level splits, which introduces leakage and optimistic accuracy^3–6. Labels for viral versus bacterial pneumonia are often weak or noisy, and class imbalance is substantial, leading to biased decision thresholds and inflated accuracy relative to macro-averaged metrics. Most non-Bayesian models report point estimates without calibrated uncertainty and show over-confidence under distribution shift, while segmentation is rarely used to suppress non-pulmonary confounders (text markers, devices). Robustness to common degradations (noise, blur, compression, downsampling) is seldom quantified, and external multi-site validation is infrequent⁷.

As of May 28, 2021, COVID-19 has affected 5.5 million individuals across 216 regions, resulting in over 353,000 fatalities. Furthermore, there are still cases are recorded in the entire world, particularly in China. Therefore, moste sensitive and less time cosuming procedure are necessary to detect the While most cases are mild, severe infections are a concern. Effective patient screening is vital to halt transmission. Traditional genetic testing, while reliable, isn’t always practical. In addition, the conventional CNNs^8–11 compute point estimates and ignore parameter uncertainty, which often yields over-confident predictions in medical tasks with limited data. Bayesian deep learning applies probabilistic inference to network weights, generating predictive distributions that capture both epistemic and aleatoric uncertainty. The posterior marginalization improves model calibration, stabilizes learning on heterogeneous imaging sources, and minimizes bias arising from small or uneven data partitions. In clinical screening scenarios, these properties provide statistically interpretable confidence measures that improve the reliability of automated COVID-19 diagnosis. Chest imaging, like X-rays or CT scans, is a supplementary screening method.

This study employs chest radiography to identify COVID-19 indicators, addressing challenges in distinguishing it from other pneumonias. Convolutional Neural Networks (CNNs) in Deep Learning offer spatial understanding and image classification capabilities. Originally pioneered in 1998 by LeCun and team, CNNs excel in recognizing visual inputs^12–15. To enhance radiographic interpretation, numerous COVID-19 pneumonia detection algorithms leverage advanced CNN models. Yet, small datasets and limited categorization hinder their efficacy. Using the principle of transfer learning and the ResNet architecture, this study addresses COVID-19 detection using chest X-ray images. Two models are trained: one distinguishes COVID-19 pneumonia from non-COVID-19 pneumonia, while the other classifies COVID-19 pneumonia against virus-related pneumonia, bacterial pneumonia, and normal cases.

CXR is prioritized for frontline triage because it is fast to acquire, low cost, widely available in emergency and ward settings, and portable for bedside imaging, which reduces patient transport and infection-control burden. Compared with CT, CXR has far lower radiation dose and shorter throughput time, enabling repeat imaging during admission or surge conditions when scanner capacity is constrained. Compared with RT-PCR, CXR can provide immediate pulmonary status when molecular results are delayed or inconclusive and helps assess severity and complications such as superimposed bacterial pneumonia. CXR has lower sensitivity for early or mild disease, labeling is susceptible to noise, and radiographic patterns overlap across viral and bacterial etiologies, so imaging should be interpreted with clinical data rather than used as a standalone diagnostic. This study targets the operational role of CXR by improving calibrated, segmentation-guided classification and reporting class-wise metrics under realistic constraints. This work includes the following research questions.

RQ1. Can a segmentation-guided Bayesian ensemble improve four-class CXR classification under patient-wise splits compared with single non-Bayesian baselines

RQ2. Does segmentation reduce spurious cues and improve separation of viral versus bacterial pneumonia

RQ3. Are predictions calibrated and reliable under common image degradations such as noise, blur, compression, and downsampling

RQ4. Can the system provide clinically usable severity grading for COVID-19 cases from CXR

Related work

In recent research¹⁶, convolution layers and capsules were employed to address class imbalance issues in Covid-19 detection. The study revealed the effectiveness of COVIDCAPS in achieving high performance with minimal trainable parameters. The model’s public availability on Github was noted, demonstrating 95.80% specificity with fewer parameters than existing models. Investigating initial Covid-19 cases in France¹⁷, two cases were identified in Paris, and the third in Bordeaux, with the patients having traveled from Wuhan. A technique based on chest X-ray images was proposed¹⁸ for Covid-19 identification. A MERS radiologic evaluation focused on a 30-year-old man with new coronavirus infection, presenting symptoms of diarrhea, fever, and stomach discomfort. The use of chest X-rays in treating unwell patients was discussed¹⁹. Employing an SVM approach, researchers studied pneumothorax diagnosis²⁰, utilizing Local Binary Pattern for lung image feature extraction. Multi-scale texture segmentation and Sobel Edge detection aided in distinguishing impaired lung regions²¹. Moreover, researchers emphasized the value of CXR in identifying lung issues²². A Covid-19 dataset was used to develop a COVID-RENet model²³, combining convolutional neural networks (CNN) and support vector machine (SVM) for enhanced classification. Similar approaches aided the detection of Covid-19’s effects on pneumonia²⁴. Renal implications were explored²⁵, with viral and serological shedding kinetics studied and a link found between lung infections and recovery²⁶. A global perspective on Covid-19 cases and fatalities was provided²⁷, suggesting X-ray scans for quick diagnosis²⁸, and Covid-19 presence in X-ray images was demonstrated²⁹. Challenges in Covid-19 detection were addressed³⁰, advocating for automated systems to prevent transmission. Difficulty distinguishing Covid-19-caused pneumonia from other symptoms was highlighted³¹. Researchers further examined the use of AI on limited X-ray images³², achieving high accuracy with deep learning and transfer learning³³. A KE Sieve Neural Network for Covid-19 analysis using Chest X-ray images was proposed³⁴, achieving 98% accuracy. A CNN-based approach was developed for pneumonia analysis³⁵, with SVM boosting accuracy. Deep anomaly detection methods³⁶ improved patient screening, detecting 70 out of 100 Covid-19 positive cases. Effects of Covid-19 on humans were studied³⁵, investigating pneumonia-associated CT scans. The potential impact of commercial travel on Covid-19 spread was considered³⁷. A hybrid system using CNN and softmax classifier was suggested³⁸ for Covid-19 detection. Radiologic research³⁹ on a novel coronavirus associated with MERS was discussed, along with the use of chest X-rays²⁷. Chest X-rays and CT scans were analyzed⁴⁰, aiming for enhanced results. Segmentation-only resources⁴¹ enable consistent masks but stop short of end-to-end evaluation on multi-class COVID-19 tasks. Classification-only studies^42,43 report strong accuracy yet often omit patient-wise splits, macro metrics, and uncertainty. Segmentation→classification approaches⁴⁴ improve specificity but seldom quantify calibration or robustness under noise/blur/compression/downsampling. Transformer-focused works⁴⁵ advance attention mechanisms but are not yet routinely evaluated for calibration and stress robustness in small-sample CXR settings. Our work combines segmentation-guided inputs, Bayesian ensembling, severity grading, calibrated decision-making, and a standardized corruption suite, and it reports class-wise metrics, macro AUC, kappa, and expected calibration error on a held-out test split.

Recent contributions have demonstrated substantial progress in automated infectious-disease diagnosis using hybrid and deep-learning architectures.⁹ developed and compared multiple machine-learning and deep-learning algorithms for COVID-19 detection from laboratory data, evaluating KNN, Decision Tree, Naïve Bayes, DNN, CNN, and LSTM models. A fused deep-learning approach is presented in¹⁰ for joint discrimination of COVID-19, tuberculosis, and normal lung conditions using chest X-ray imaging. Their model integrated Crimmins speckle filtering, generative adversarial data augmentation, and feature fusion between DenseNet121 and ResNet50 backbones, followed by bidirectional long short-term memory (Bi-LSTM) classification. In another study,¹¹ examined the impact of image-filtering and feature-extraction techniques on COVID-19 classification using radiography images.

Major contributions

1. Using advanced CNNs, the study pioneers sophisticated deep learning techniques to enhance the analysis of medical images, particularly chest X-rays.

2. Recognizing the pressing need for rapid COVID-19 screening without a vaccine, the research harnesses the potential of chest X-rays as a diagnostic tool to swiftly identify potential cases, thereby aiding in timely intervention.

3. Thoroughly investigating an array of deep learning models, such as VGG-16, RestNet-34, InceptionV3, VGG-19, and ResNet-50, the study enhances the accuracy of diagnostic outcomes. Establishing ResNet-50 as a gold standard benchmark model for robust evaluation and performance comparison of other deep learning models.

4. Pioneering the development of an automated system, the research achieves effective differentiation between chest X-ray images depicting COVID-19 cases and non-COVID-19 conditions. This novel automation contributes to the acceleration and efficiency of the screening process, aiding medical professionals in making accurate and prompt diagnostic decisions.

Convolutional neural network

The CNN in CXR analysis must cope with small, imbalanced cohorts, label noise for Spneumonia subtypes, and domain shift across devices and sites. They are also vulnerable to spurious non-pulmonary cues (text markers, devices) and over-confidence under distribution changes. Our design targets these issues as follows. A ResNet-34 backbone (224×224 input, ImageNet initialization) is fine-tuned end-to-end with patient-wise splits. A lung/opacity segmentation branch generates masks that gate the classifier’s inputs and features, suppressing non-pulmonary regions and emphasizing parenchymal disease patterns. We train K diverse classifier instances (different seeds/folds/backbone initializations) and aggregate logits with temperature scaling learned on a validation split, which improves calibration and minority-class recall. Class imbalance is handled with weighted cross-entropy; sensitivity analyses with focal loss were consistent. Data augmentation mirrors radiographic variation (horizontal flip, small rotations, scale/contrast jitter, mild blur/compression) to improve robustness. We report class-wise precision/recall/F1, macro AUC, Cohen’s kappa, and expected calibration error, and we quantify robustness with a corruption stress test (noise, blur, compression, downsampling) on the held-out test set. This task-focused configuration is more informative for medical imaging than generic CNN descriptions and directly addresses the known failure modes in CXR classification. Hyperparameters are detailed in Table 1, with the CNN architecture utilizing a 244x244 breast image tensor featuring five layers and convolution layers with 64 filters and a stride of 1x1 in the first layer.

Table 1.

Several strengthened limits of the CNN model construction and favored weights are investigated in work.

M1	M2	M3	M4	M5	M6	M7	M8
CNN	224 × 224	RMS prop	32	10 fold	3e 4	BCE CCE	50

A 2 x 2 stride max-pooling layer 112 x112 takes the output from the last layer and decreases the original size. The outcome of the pooling layer is distributed throughout the system using the ReLU start function. The nonlinear effect is now passed on to the following convolution layer, which has the same 5 x 5 x 64 filters as the previous one and a stride value of 1 x 1. The nonlinear output is now passed on to the next convolution layer, which has the same 5 x 5 x 64 filters as the previous one and a stride value of 1 x 1 in the next step, resulting in an output that is half the size of the original 56 x 56 input. Following the activation of the ReLU, the production of the third convolution layer is transmitted. The output is routed via a max-pooling layer, resulting in type 28 x 28 tensors as a consequence of the routing. The work is then transferred to the fourth convolution layer, which has the same stride as the other three levels and 512 filters with a kernel size of 5 x 5 x 256. The fourth convolution layer has the same stride as the previous three layers and 512 filters with a kernel size of 5 x 5 x 256.

It is necessary to activate a fifth convolution level that has 512 screens, and an array of 14 x 14 x 512 kernels to get the desired result; the output of that layer is max-pooled with a size 2 x 2 stride to provide a size 14 x 14 output, and the work of that layer is max-pooled with a size 2 x 2 stride to provide a size 14 x 14 output when that layer is triggered. The tensor produced has the form 7 x 7 x 512, a prime number. The resulting tensor is flattened when there are 25,088 neurons. When neurons arise, the weighted values that emerge are comparable to the symptoms of COVID-19, and the dropout layer is utilized to reject input to prevent the overfitting of the network. For the sake of my study, this research used a 0.5 training dropout rate. As a result of the connected layer, the tensor is reduced from 25,088 to 64 neurons, with ReLU activation included in the output. The entirely connected layers offer a tensor of 64 neurons, which are transformed into neuron counts based on how many categories the retinal image matches: healthy, COVID-19, and pneumonia, depending on how many layers are present.

Convolution layer

This layer is made up of a number of different filters (kernel). After each filter is applied to the input picture, features are recovered by creating a new layer on top of the existing one. Each layer represents one or more of the most vital components or physiognomies of the underlying image. The convolution is operating⁴⁶ as

1

We have used the standard definition of the Adjoint F* of a linear map F F*t: F* is define as linear map satisfying. Assuming that t only takes integer-values⁴⁷ and can be estimated as

2

F(t) is a linear map that swaps the domain and codomain of F. The direction-reversing property is met by the Adjoint operator. In what came before, it was assumed that the convolutional process only has one dimension. The following is the definition of a two-dimensional convolution action using input in (m, n) and a kernel f (a, b).

3

Here are some of the other things With the goal of not overfitting, the measure of error and the regularity of the parameter set are used to avoid overfitting. After the learning step is done, F* can be used to put together a new image from the measurements of the old one. In accordance with the commutative rule, the kernel is inverted, and the following is identical to

4

To sum up, the reconstruction function in the objective function approach is itself a regularised minimization problem, while the solution to a regularised minimization problem in the learning approach is a parametric function that can be used to solve the inverse problem. Neural networks use a function called cross-correlation, which is similar to convolution but doesn’t require flipping the kernel like convolution does⁴⁸.

5

The values of each layer are based on the values of the layers that came before it. The next layers are given by where F is the training data that was put in, t is the weight, When this layer sets a negative input value to zero, it speeds up and improves analyses and training and keeps the gradient from disappearing completely. It can be described as

6

Here, you’ll find a list of CNN’s most popular design models. Also, I’ll list architecture models with their features and comparisons to other models in order of when they were made. I’ll also talk about the theories behind each architecture improvement. But it’s hard to figure out how to draw some mathematical rules because all well-known architectures were made for a specific problem, and few papers talk about the mathematical theory of CNN, in which the neuron gets the input x⁴⁹.

Pooling-max layer

This Layer employs a model-based discretization strategy to solve the problem. It is used to develop assumptions about which components may be accessible in the binned sub-regions of the input design and to reduce the dimensionality of the input design, both of which are beneficial (input picture, hidden layers, output matrix, and so on). Even if the number of learning parameters is reduced, there is still a significant amount of internal depiction interpretation invariance kept. During the Carpooling method, my model employed a kernel dimension of 3 x 3 during the pooling max procedure. Following the completion of the last convolution block, the network is reduced to a single dimension⁵⁰.

7

I used multiple convolution filters or kernels that run over the image and compute a dot product. Each filter extracts different features from the image. Let’s consider a filter of size 2x2 and an image of size 5x5. I will perform an element wise multiplication between the image pixel values that match the size of the kernel and the kernel itself and sum them up. This provides us a single value for the feature cell⁵¹.

Loss function

When the class score vector is composed of j elements, y signifies the CNN score for the j^th member of the positive class, and c denotes the total number of classes for each picture. In the logarithmic form of the equation, the SoftMax provides a reasonable prediction probability.

8

The hyperparameters are changed so that the average loss is as small as possible. I found the weights (L_i) and biases (e^by) that make the value of J as small as possible (average loss). This reminds me of residuals, which are used in statistics to measure how far the actual y values are from the regression line (predicted values). The goal is to make the net distance as short as possible.

Accuracy of regularization

In this case, a dropout is employed as a reliable regularisation process since it is easy to implement. In our investigation, the hyper parameter is set to 0.50 since it produced the greatest regularisation results.

Proposed methodology

This research involves three experiments, each designed to test the models and the effects of different steps. Each experiment uses a different dataset. The same images are used for all COVID-19 positive cases, but three different datasets simulate challenging scenarios. Experiments 1 and 2 use positive and negative case datasets separately, while Experiment 3 uses images from the COVID-19 imaging system (images from 2019-2020). The chest X-ray images are from two sources, comprising 100 locations. The COVID-19 images are from the Cohen JP’s dataset , which combines images from various sources and is updated regularly. The normal images are from GitHub repository 1 , which has 500 images from the ChestX-ray dataset . The dataset is classified into a 71% training subset and a 29% testing subset, due to limited information on data partitioning in the datasets. To reduce discrepancies from different image projections (frontal and lateral) in the COVID-19 datasets, a filtering technique is used. It keeps frontal projection images, which have more data than lateral views. This technique filters the COVID-19 dataset. The research has different phases, including data acquisition. The datasets are collected from other sources, such as GitHub and Kaggle repositories. GitHub repositories are used for covid images, and Kaggle repositories are used for normal images. Then, the dataset is categorized into training and validation sets in a 70:30 ratio. Next, there is pre-processing and training. Due to the limited number of images in the dataset, online augmentation is performed during each epoch and applied transfer learning and fine-tuning to my model. ResNet50 is used to design this model, because it has an optimal number of layers in the network. The validation accuracy of the trained model are justified in Fig 5, which is discussed in detail later in this paper.

The experimental diagram (Fig 1) shows the steps: ‘a’ is data arrangement, ‘b’ is lung segmentation, ‘c’ is COVID-19 prediction using normal images, ‘d’ is COVID-19 prediction using only the lung segment, and ‘e’ is COVID-19 prediction using only the lung segment. Fig 1 shows how raw X-rays are converted into positive and negative results, which are used for the final results. It converts raw X-rays into informative images and gives me the result of the positive and negative information images.

Fig. 1.

Converting x-ray to positive and negative result.

The images come from different datasets with different sizes and conditions, as they are taken from various X-ray machines in Pakistan, Japan, and other countries. To reduce the data variability, a preprocessing image phase is used with three steps: (1) Resize all images in one channel to 224 x 224 pixels, (2) Resize all images in another channel to 224 x 224 pixels and (3) Resize all images in a third channel (grayscale) to 224 x 224 pixels.

Then, the datasets is normalized using Eq. (9), where x is the original image and N is the normalized image. Finally, the datasets is standardized using Eq. (10), where Z is the standardized image and N is the normalized image ⁵².

9

10

Several approaches are used to segment images based on the U-Net architecture. It is previously discovered that the U-Net architecture is successful in segmenting medical pictures in a previous research study. This kind of model aims to generate an image mask that contains one (1) in the reconstruction region and zeros (0) everywhere else. Therefore, in a production situation, the model’s input is an X-ray chest picture, and the model’s output is a predicted mask of the chest image. To determine the optimal number of filters for this research, have three alternative numbers of layers on CNN at various times. Calculating the number of screens in contraction blocks is done using Eq. (11), where F is the number of initial filters, and i is the number of contraction blocks using the formula. Expansion blocks are represented by Eq. (12), which provides information on the number of filters in each block. While the number of filters in the preceding contraction block is represented by the letter F, and the number of filters in the following expansion block is represented by the letter I, respectively. The number of filters in the transposed convolution layer is the same as the number of filters in the convolutional layers of the expansion block ⁵³.

11

12

Therefore, in a production situation, the model’s input is an X-ray chest picture, and the model’s output is a predicted mask of the chest image.

Convolutional layers with the standard kernel initialization and padding have kernel sizes of 3 x 3 with the normal initialization and padding. When it comes to the first two blocks of Expansion and contraction, the pool size is 2 x 2, and the dropout rate is 0.1 in 0.2 in the third and fourth blocks and 0.3 in the fifth. During the first two Expansion and contraction blocks, there is no dropout. In transposed convolutional layers, the kernel size is 2, the strides are 2, and the padding is the same as in the original convolutional layers. When it comes to the final convolutional layer, a single filter with a kernel size of 1 x 1 is applied.

According to the ILSVRC 2015 classification competition results, this model is very successful. Its ensemble won first place with an error rate of 3.57 percent, indicating that it is highly successful. It also brought first-place honors in the ILSVRC and COCO contests in 2015 for ImageNet detection, imagine localization, and COCO detection & segmentation, among other accomplishments. Because deep CNNs are exceptionally good at recognizing low, mid, and high-level attributes from images and stacking more layers is commonly associated with improved accuracy, the question arises: can a model performance be improved simply by stacking more layers? Multiple ways have been devised to cope with the problem of vanishing/exploding gradients, resulting in networks with tens of layers capable of convergent behavior. A new problem occurs when deep NNs converge; however: accuracy becomes saturated and decreases quickly. In addition, more layers to an appropriate deep model only raised the training error. An initial shallow model is used to create a deep learning model, which is then used to create a shallower model. As shown in Figure 2 that the deeper model is constantly adding more error, which is something that shouldn’t be happening^8,54.

Fig. 2.

Training and test error graphs.

It is possible for the layers to fit a residual mapping, H(x), and for the non-linear layers to fit another mapping, F(x):=H(x)x, as indicated in Fig 3. This resulted in H(x):=F(x)+x. Quick identity mapping has the advantage of not requiring any more parameters to be added to the model, as well as reducing calculation time. Table 2 represents the ResNet 50 architecture.

Fig. 3.

Resnet mathematical equation and model.

Table 2.

ResNet50 architecture and 18, 34, 50, 101, 152 layers models.

Name layer	Size output	Fill 1	Fill 2	Fill 3	Fill 4	Fill 5
conv1	112 112	7 7, 64, Stride 2
conv2.	56 56	3 3max pool, stride2
conv3.	28 28
conv4.	14 14
conv5.	7
	1 1	Normal pool, 1000-d fc, softmax
FLOPs		1.8 109	3.6 109	3.8 109	7.6 109	11.3 109

Because the available dataset contained only a few hundred chest X-ray samples, several measures were adopted to prevent overfitting and enhance generalization. The training pipeline integrated real-time augmentation operations such as rotation, scaling, and flipping to generate diverse views of each image. Model optimization employed dropout, batch normalization, and weight decay to regularize learning. Early stopping and five-fold cross-validation were applied to monitor validation loss and provide unbiased performance estimation. The Bayesian inference layer further reduced overfitting by performing posterior sampling over weights rather than deterministic fitting, ensuring that the final model generalized well within the constrained data regime. The original dataset presented a significant imbalance between COVID-19-positive and non-COVID-19 samples. To alleviate class bias, stratified sampling was employed during data partitioning, and the loss function was reweighted in proportion to inverse class frequency. Augmentation operations rotation, reflection, and contrast adjustments were applied more intensively to the minority class to improve sample diversity. The model’s effectiveness was evaluated through class-wise precision, recall, and F1-scores, confirming balanced performance across categories. The Bayesian framework’s uncertainty quantification further minimized prediction bias by down-weighting uncertain outputs and reducing the model’s dependence on majority-class patterns.

To reduce variance and bias arising from the small dataset, we used patient-level, stratified 5-fold cross-validation. Class imbalance was handled with inverse-frequency weights in the cross-entropy objective and minority-focused augmentation (rotation ±10°, horizontal flip p=0.5, scale 0.9–1.1, contrast normalization). The classifier was trained with dropout 0.5, batch normalization, and L2 weight decay 1e−4; early stopping monitored validation loss with patience 10. Bayesian inference was implemented via Monte Carlo dropout at test time (T=30) to obtain predictive distributions and uncertainty estimates.

Proposed framework

Figure 4 presents the porposed framework where two different datasets are used to train the models. The first model is trained on a dataset that had just normal and pneumonia patients. After learning from the first model, the second model is prepared using a dataset that comprised COVID-19 patients as well as patients who had pneumonia and were otherwise healthy. The goal of this example is to demonstrate how transfer learning had an impact on the final design of the proposed model.

Fig. 4.

Proposed model of evaluating the ResNet and Xpection.

Results and discussion

The dataset for this research is taken from the Kaggle repository. It consists of chest X-ray images of Covid-19 patients who were average. This dataset is not meant to claim the diagnostic ability of any deep learning model, but to be used for research on various methods of detecting Coronavirus infections using computer vision techniques. The dataset has 1531 images, which are divided into training (5467), validation (965), and covid sets. The validation set has 238 normal, 86 covid, and 641 pneumonia samples. For this work firstly, 576 covid scans collected from a posterior-anterior (PA) view and resized them to 128 by 128. These scans are used for our covid set based on the the data distribution for training and testing.

The research has different phases, firstly, there is data acquisition. The datasets is collected from other sources, such as GitHub and Kaggle repositories. GitHub repositories are used for covid images, and Kaggle repositories are used for normal images. Then, the dataset is classified into training and validation sets in a 70:30 ratio as mentioned in Table 3. Figure 5 shows the validation accuracy of the trained model at 98.7%. The infected chest X-ray images are labeled as positive or negative cases.

Table 3.

Classification results comparison of all eight CNN models.

Model	Pre (%)	Sen (%)	Spe (%)	F1-score	Acc (%)	AUC
ResNet-50	96.77	100.00	96.67	0.9836	98.33	0.9836
ResNet-34	95.24	100.00	95.00	0.9756	97.50	0.9731
VGG-16	95.08	96.67	95.00	0.9587	95.83	0.9487
VGG-19	98.24	93.33	98.33	0.9573	95.83	0.9506
Inception-V3	96.36	88.33	96.67	0.9217	92.50	0.9342

Fig. 5.

Validation accuracy of trained network model.

Then created the GUI for MATLAB for detection individual x-ray. For negative cases and as well as for positive cases. This help in the individual cases for scanning and detecting the cases. Covid-19 positive case are shown in Fig. 6 and negative case individually are shown in Fig. 7.

Fig. 6.

Individual result for covid-19 positive case.

Fig. 7.

Individual result for covid-19 negative case.

Figure 8 Indicates the spooling of chest x-rays and separating the infected images and non-infected images by implementing the ResNet50 algorithm.

Fig. 8.

ResNet50 spooling x-ray images.

Figure 9 consists of the spooling of chest x-rays and separating the infected images and non-infected images by implementing the ResNet50 algorithm.

Fig. 9.

Spooling image of ResNet50.

Figure 10 shows the final result of the simulation model and as the output of the result the lost data is as low rate as previous and the accuracy rate is high from the previous rate and this is the final required result of the proposed framwork.

Fig. 10.

Final training progress epoch 5 out of 5.

Figure 11 declares the final result of chest X-ray images with infected as darkest and the lightest as negative images of the result.

Fig. 11.

Final result ResNet50.

The detailed categorization and comparison report are listed in Table 4 and its in ascending order as the lowest on bottom and newest on top. Inception V3 accuracy 92.50, VGG-19 Accuracy 95.83, VGG-16 Accuracy rate is 95.83 same as the updated model but ResNet-34 Accuracy rate is 97.60 and we have the precision and accuracy rate at top with ResNet-50 model with 98.33%. Similarly Tables 5, 6 and 7 declare the processing summary, model summary and correlation of the proposed model. VGG19 FL refers to a VGG19 backbone with a frozen feature extractor (convolutional blocks fixed) and a lightweight task head fine-tuned on our CXR data, used as a baseline without Bayesian ensembling.

Table 4.

Sequence plot.

Model description
Model name		MOD_1
Series or sequence	1	Confirmed_mean
	2	Deaths_mean
	3	Recovered_mean
Transformation		None
Non-seasonal differencing		0
Seasonal differencing		0
Length of seasonal period		No periodicity
Horizontal axis labels		ObservationDate
Intervention onsets		None
For each observation		Values not joined

Table 5.

Case processing summary.

Case processing summary
		Confirmed_mean	Deaths_mean	Recovered_mean
Series or sequence length		494	494	494
Number of missing values in the Plot	User-Missing	0	0	0
	System-Missing	0	0	0

Table 6.

Model summary.

Normal_count
Frequency			Percent	Valid Percent	Cumulative Percent
Valid	0.22	1	0.1	6.3	6.3
	2.57	1	0.1	6.3	12.5
	2.85	1	0.1	6.3	18.8
	2.88	1	0.1	6.3	25.0
	3.13	1	0.1	6.3	31.3
	3.15	1	0.1	6.3	37.5
	3.98	1	0.1	6.3	43.8
	3.99	1	0.1	6.3	50.0
	5.50	1	0.1	6.3	56.3
	6.37	1	0.1	6.3	62.5
	6.40	1	0.1	6.3	68.8
	6.84	1	0.1	6.3	75.0
	6.91	1	0.1	6.3	81.3
	7.00	1	0.1	6.3	87.5
	7.40	1	0.1	6.3	93.8
	11.20	1	0.1	6.3	100.0
	Total	16	1.7	100.0
Missing	System	934	98.3
Total		950	100.0

Table 7.

Correlation

Correlations
		Confirmed_mean	Deaths_mean	Recovered_mean	GDPpercapita
Confirmed_mean	Pearson correlation	1	0.889**	0.986**	0.234**
	Sig. (2-tailed)		0.000	0.000	0.004
	N	150	150	150	150
Deaths_mean	Pearson correlation	0.889**	1	0.861**	0.187*
	Sig. (2-tailed)	0.000		0.000	0.022
	N	150	150	150	150
Recovered_mean	Pearson correlation	0.986**	0.861**	1	0.198*
	Sig. (2-tailed)	0.000	0.000		0.015
	N	150	150	150	150
GDPpercapita	Pearson correlation	0.234**	0.187*	0.198*	1
	Sig. (2-tailed)	0.004	0.022	0.015
	N	150	150	150	150

To obtain class-robust evaluation under imbalance, we computed class-wise ROC curves and AUC together with per-class precision, recall, and F1, their macro-averages, and Cohen’s Kappa. Figure 12 shows that the detector maintains stable discrimination across all classes with area under curve (AUC) values of 0.808 (COVID-19), 0.824 (Bacterial pneumonia), 0.793 (Viral pneumonia), and 0.766 (Normal), yielding a macro AUC of 0.798. These results confirm balanced discrimination rather than majority-class dominance. The full set of per-class precision, recall, F1, macro-averages, and Cohen’s Kappa are reported in Table 8.

Fig. 12.

Reciever operating features per class on the test set with AUC.

Table 8.

Class-wise and macro-averaged performance on the test set.

Class	Precision	Recall	F1	AUC	Support
COVID-19	0.540	0.540	0.540	0.808	50
Bacterial pneumonia	0.564	0.554	0.559	0.824	56
Viral pneumonia	0.522	0.533	0.528	0.793	45
Normal	0.571	0.571	0.571	0.766	49
Macro average	0.549	0.550	0.549	0.798	–
Cohen’s Kappa	–	–	–	0.3990	–

Table 9 shows that our study reports fivefold, patient-level cross-validation on 1,531 CXRs with accuracy 98.33 percent and macro-AUC 0.798, alongside class-aware metrics and calibration analysis. References ¹⁰ and ¹¹ also report ≥ 98 percent accuracy on internally split CXR datasets, while reference ⁹ uses non-imaging laboratory data and is not directly comparable to imaging performance. None of the cited works report multi-institutional external validation in the abstracts, so all headline figures, including ours, should be interpreted as internal performance.

Table 9.

Comparison of our cross-validated , including modality, task, dataset size, validation, metrics, headline result, and external validation status.

Source	Modality	Task / Classes	Dataset size	Validation	Metrics reported	Headline result	External validation
Our study	CXR	COVID-19 vs Pneumonia vs Normal (3-class)	1,531 images	fivefold stratified CV (patient-level)	Accuracy, macro AUC, macro F1, Cohen’s Kappa	Accuracy 98.33%; macro AUC 0.798	Planned (COVIDx, BIMCV-COVID19 +)
9	Laboratory data (not imaging)	Binary COVID-19 vs non-COVID	5,644 lab records (10% positive)	Single split (internal)	Accuracy, F1, Precision, Recall	LSTM: Acc 96.78%; F1 96.58	No
10	CXR	COVID-19 vs TB vs Normal (3-class)	Not specified in abstract (modest)	Hold-out (internal)	Accuracy (plus method description)	Accuracy 98.22	No
11	CXR	Binary COVID-19 vs Normal	Not specified in abstract (small–moderate)	Hold-out (internal)	Accuracy (plus feature extraction timing)	Accuracy up to 99.93 (LDA + SVM)	No

We observed that subtype discrimination is the main error mode. The confusion matrix in Figure 13 shows reciprocal misclassification between bacterial and viral pneumonia of 0.20 each, while leakage into COVID-19 or normal is smaller (viral→COVID-19 0.13, viral→normal 0.13; bacterial→COVID-19 0.12, bacterial→normal 0.12). These patterns are consistent with overlapping radiographic phenotypes and label noise in legacy datasets. To mitigate this, we applied lung-field masking to suppress background confounders, inverse-frequency class weights in the loss, targeted augmentation to increase texture diversity, and Monte Carlo dropout at test time to flag high-variance, low-confidence subtype predictions for human review. Subtype AUCs and F1-scores reported in Table 8 (bacterial AUC 0.824, F1 0.559; viral AUC 0.793, F1 0.528) align with the confusion trends.

Fig. 13.

Normalized confusion matrix on the test set.

Robustness was quantified by re-evaluating the trained model on corrupted copies of the test images as discussed in Figure 14. We applied Gaussian noise (sigma 5,10,15,20,25), motion blur (length 5,9,13,17,21), defocus blur (radius 1,2,3,4,5), JPEG compression (quality 95,85,70,50,30), and downsampling followed by bicubic upsampling (448,384,320,224,128). Metrics reported per severity were macro AUC, macro F1, and expected calibration error using 15 bins. The model weights and thresholds remained fixed throughout.

Fig. 14.

Robustness to common degradations on the test set: macro AUC versus ordinal severity.

Discussion

Numerous studies have emerged following the release of the Cohen dataset³⁴, aiming to predict COVID-19 infection with considerable accuracy through chest X-ray images. Consequently, a substantial effort has been devoted to constructing practical diagnostic models using deep learning (DL) methodologies. While CNN-based networks have incorporated the concept of time lag (TL) over the years, many previous strategies were evaluated on limited data points, resulting in erroneous information.

To evaluate the performance of various neural networks in predicting COVID-19 infections, X-ray images are employed. The networks assessed include AlexNet, VGG-16, MobileNet-V2, SqueezeNet, ResNet-34, and Inception-V3. Rigorous testing on a substantial dataset, considering multiple factors, leads to the identification of the most proficient model for automated COVID-19 screening. Two distinct sources of COVID-19 and normal X-ray images, namely³⁴ and³⁵, are utilized to address data imbalance concerns. Through experimental data and comprehensive analysis, ResNet models prove superior. Leveraging the ResNet-34 network achieves accuracy and sensitivity rates of 98.33% and 100.00%, respectively. Importantly, this model carries no additional costs and has the potential to aid radiologists in confirming their findings. An essential aim of this study is to enhance real-time patient quarantine decisions, thereby curbing the spread of COVID-19. However, a notable limitation is the evaluation’s reliance on a small number of COVID-19 samples due to the pandemic’s early stage and limited publicly available datasets. The next phase involves testing the proposed technique on extensive datasets in the near future.

The trained model was evaluated using internal cross-validation due to limited access to large, institutionally diverse datasets at the time of experimentation. Although the internal validation results indicate strong diagnostic performance, external validation remains essential for assessing domain robustness and clinical transferability. Future work will incorporate additional publicly available repositories, including COVIDx and BIMCV-COVID19+, to verify model generalization across heterogeneous imaging sources and patient demographics. The uncertainty-aware Bayesian inference employed in this study is expected to provide more stable predictive behavior under such cross-domain evaluation. While cross-validation and uncertainty-aware inference reduce variance and overfitting risk, internal resampling cannot substitute for external validation. We therefore plan testing on independent multi-institutional datasets to confirm generalizability under domain shift. Planned next steps are external patient-wise and site-held-out validation on COVIDx and BIMCV-COVID19+ with macro-AUC, macro F1, Brier score, and calibration error with 95% CIs, plus targeted expansion of viral and bacterial cases with adjudicated labels. We will also pilot CT late-fusion and a self-supervised CXR backbone; success is defined as delta AUC ≥ 0.02 or improved calibration at matched sensitivity without increasing false negatives.

Conclusion

This study unveils a DL-based automated approach for effectively distinguishing COVID-19-infected individuals from healthy ones using chest X-ray images. Various pre-trained CNN architectures, grounded in the TL concept, are subjected to testing using publicly accessible X-ray data, and their outcomes are compared. ResNet-34 stands out with an impressive accuracy rate of 98.33%, surpassing competing networks in potential for predicting COVID-19 infection. Radiologists can employ this model to verify their screening outcomes efficiently, benefiting both parties. Furthermore, this breakthrough paves the way for the development of more efficient CNN models based on residual connections for enhanced COVID-19 detection. While limited research has explored multi-class classification, the proposed DL model primarily targets high performance in binary classification (COVID-19 vs. normal) and holds potential for extended application (COVID-19 vs. pneumonia vs. normal). Future research focuses on assessing the model’s effectiveness in multi-class classification and exploring optimization techniques to enhance model reliability.

Author contributions

Conceptualization: Yuanyuan Chen, Waqas Khan, Farman Ali Methodology: Yuanyuan Chen, Farman Ali, Haleem Afsar Software: Waqas Khan, Farman Ali Validation: Waqas Khan, Munaza Ijaz Formal analysis: Farman Ali, Yuanyuan Chen Investigation: Haleem Afsar, Munaza Ijaz Resources: Ali Alshamrani, Shahid Kamal Data curation: Waqas Khan, Munaza Ijaz Writing – original draft: Yuanyuan Chen, Farman Ali Writing – review and editing: Yuanyuan Chen, Waqas Khan, Farman Ali, Haleem Afsar, Munaza Ijaz, Ali Alshamrani, Shahid Kamal Visualization: Waqas Khan, Farman Ali Supervision: Ali Alshamrani, Shahid Kamal Project administration: Yuanyuan Chen, Shahid Kamal Funding acquisition: Ali Alshamrani, Shahid Kamal All authors reviewed and approved the final manuscript.

Funding

"This research is supported by Multimedia University (MMU) through its Article Page Charge (APC) Sponsorship Scheme." And “This research was funded by Taif University, Saudi Arabia, Project No.Funding declaration:(TU-DSPP-2025-09)."

Data availability

Data is provided within the manuscript or supplementary information files.

Declarations

Competing interests

The authors declare no competing interests.