Detection of bacterial and viral pneumonia in pediatric chest radiographs using fusion of mediastinum and lung imaging biomarkers

1 Introduction

Pneumonia is a fatal disease that restricts the reception of oxygen-rich air and increases difficulty in breathing (Tulo et al., 2023a). Children with comorbidities such as asthma, diabetes, and pulmonary problems are prone to the disease. According to the World Health Organization (WHO), around two million people lose their lives to pneumonia each year, with the majority being children younger than five (Rajaraman et al., 2018). Bacterial and viral pathogens are the two primary causes of pneumonia and need different forms of treatment.

Early diagnosis is essential to enable timely patient referrals. Traditionally, pneumonia is diagnosed through assessment of symptoms, Chest Radiograph (CXRs), blood examinations, and sputum culture analysis. Conventional blood and culture tests are constrained by delayed results and are resource-intensive, while CXRs offer a rapid and widely accessible modality. CXR is the recommended radiological examination for screening and diagnosis of chest abnormalities (Rajaraman et al., 2018; Tulo et al., 2023a).

CXRs offer benefits such as being non-invasive, involving minimal radiation exposure, portable in nature, and cost-effective (World Health Organization, 2016). Proper analysis of these images can indicate the existence of potential abnormalities (O’Grady et al., 2014). However, the manual interpretation of these images is reported to be complex and is limited by factors such as radiologists’ heavy workload and variable experience. Hence, a fast and accurate computerized system is required to analyze these images.

Computer-Aided Diagnosis (CAD) systems combining image processing and Machine Learning (ML) techniques could help in the automated analysis of images to detect the disease (Garg et al., 2019). These systems can assist radiologists and clinicians in the preliminary diagnosis of chest-related illnesses, including pneumonia (Wang et al., 2021; Isaac et al., 2022). These can aid in the triage and follow-up of the cases that need immediate attention from the experts. The performance of CAD systems mainly depends on the segmentation, biomarker extraction, and classification (Rajaraman et al., 2018).

In CXR images, the anatomical regions such as lungs and mediastinum are considered clinically significant for diagnosing chest-related abnormalities (Thacker et al., 2015; Tulo et al., 2023b). Mediastinum is the thoracic cavity’s central compartment, laterally bounded by the lungs. Pneumonia images exhibit manifestations such as patchy areas of consolidation, bronchial wall thickening, enlarged hilar lymph nodes, and pleural effusion (O’Grady et al., 2014; Garg et al., 2019). Bronchial dilatation enlarges the airways, while mediastinal lymphadenopathy causes lymph node enlargement, increasing the width of the mediastinum (Garg et al., 2019). Pleural effusion shifts the superior and inferior mediastinum due to the fluid accumulated in the pleural space (Khajotia, 2012). These abnormalities result in structural alterations in both the lungs and the mediastinum. The radiological signs associated with these regions aid in detecting pneumonia conditions (Whitten et al., 2007). Hence, analysis of structural alterations in these regions using CAD systems could contribute to better diagnostic outcomes (Whitten et al., 2007).

Segmentation of the lungs and the mediastinum from CXR images is essential in developing the CAD system (Li et al., 2024). In literature, Volpi et al. (2009) used a snake operator, Kumar et al. (2020) employed a Convolutional Neural Network (CNN), and Zhang et al. (2013) explored the context-contour method to segment the regions in CT images. Liu et al. (2016) initially segmented the lungs using thresholding and used the delineated lungs and convex hull to segment the mediastinum. Recently, Tulo et al. (2021) used the level set methods to segment these structures in tuberculosis CXR images.

Different anatomical regions have been segmented from medical images using level set approaches (Govindarajan and Swaminathan, 2019). These methods are adaptable to different image resolutions and can achieve subpixel precision. An edge-based technique, the Reaction-Diffusion Level Set (RDLS), eradicates the necessity of re-initialization. Govindarajan and Swaminathan (2021) implemented RDLS for lung segmentation in tuberculosis images and reported improved outcomes. Chan-Vese (CV) model is used to segment different sub-anatomic regions with different mean pixel intensities. This method has been successfully implemented in radiographic images to segment lung fields (Chung et al., 2018). Tulo et al. (2022) employed this model to segment the mediastinum in drug-resistant tuberculosis images. However, existing literature indicates that approaches for segmenting both the mediastinum and lungs in pneumonia radiographic images remain limited.

Morphological biomarkers can reveal crucial details about structural changes and have been effectively utilized to replicate the intuition of experts. Karargyris et al. (2016) employed lung biomarkers for pneumonia detection and demonstrated their ability to capture significant structural information. Tulo et al. (2022), Santosh and Antani (2018), Chandra et al. (2020) have utilized these biomarkers to diagnose tuberculosis and found that they can be used to characterize the structural variability. Hu moments are another set of morphological biomarkers that achieves the property of translation invariance by using the central moment. This approach removes the influence of scaling by normalization and eliminates the effect of rotation by constructing a polynomial (Hu, 1962; Goyal and Singh, 2023). Jaeger et al. (2014) explored the lung moment biomarkers for tuberculosis screening. However, existing studies indicate that the evaluation of mediastinum morphological biomarkers for differentiating pneumonia conditions remains limited.

Biomarker selection techniques are utilized to enhance classification performance (Wang et al., 2020). By rejecting redundant biomarkers and retaining the most discriminative ones, these approaches improve the classification efficiency (Wang et al., 2015). It also improves the generalization capability of the classifiers and decreases the training time. Wrapper-based algorithms select the candidate biomarkers by training the classifiers using different learning methods. They select the optimal biomarkers based on the classification performance (Wang et al., 2020). Guan et al. (2010) used a learning-based biomarker selection method to achieve better image retrieval performance. Mohanty et al. (2020) employed a wrapper-based selection method to identify abnormal mammographic images and achieved better accuracy.

Machine Learning (ML) offers the capability to learn meaningful patterns from complex and heterogeneous data to analyze biomedical images. ML models contribute to pattern recognition, classification, and prediction by leveraging the inherent statistical relationships within the data (Litjens et al., 2017). ML approaches such as Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and Linear Discriminant Analysis (LDA) remain widely used due to their relative interpretability, simplicity, and effectiveness in small to medium-sized datasets.

SVMs widely applied for disease classification due to their capability of handling high-dimensional data and robust generalization even with small sample sizes. Recent studies show that the classifier performs better in cancer detection and neuroimaging classification when paired with optimized biomarkers (Ul Amin et al., 2024). KNN offers an effective non-parametric approach, particularly when biomarker representations are well-defined. It has been employed in early disease screening and medical image retrieval (Wang et al., 2020). Similarly, LDA has been utilized as a dimensionality reduction and classification technique, improving class separability in tumor classification and retinal disease. These ML techniques remain highly relevant over deep learning methods, especially in scenarios with limited labelled data, low computational resources, and requirements for interpretability. Their continued integration with optimized biomarker extraction pipelines can improve disease detection.

The principal objective of this work is to improve the detection of bacterial and viral pneumonia conditions in pediatric chest radiographic images using multiple wrapper-based fusion of morphological imaging biomarkers extracted from the lung and mediastinum regions in conjunction with machine learning techniques. The key contributions of the present work include investigation of two clinically significant anatomical regions, the mediastinum and the lungs for diagnosing bacterial and viral pneumonia in pediatric chest radiographic images; implementations of a hybrid segmentation model using edge and region-based level set techniques to segment both lungs and the mediastinum; and exploration of multiple wrapper-based biomarker fusion techniques to combine the optimal mediastinum and lung biomarkers to improve the bacterial and viral pneumonia classification performance.

The remainder of this paper is structured as follows. Section 2 details the materials and methods employed in this study. Section 3 presents the obtained results along with their discussion. Finally, Section 4 summarizes the conclusions of the work, followed by the list of references.

2 Materials and methods

The workflow of the current approach is demonstrated in Figure 1. The CXRs are obtained from a publicly accessible dataset (Kermany D. et al., 2018). A hybrid segmentation model is applied to segment the significant anatomical regions. The lung segmentation is carried out using the RDLS technique. In addition, the CV model is employed to segment the mediastinum. The segmented images are processed to extract geometric and Hu moment morphological biomarkers, which are statistically examined to identify structural variations associated with pneumonia. Then, the wrapper-based biomarker fusion approaches are utilized to combine the optimal biomarkers of the mediastinum and lungs. For this, two different learning algorithms, Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM), are utilized. Further, individual and fused optimal biomarker subsets are fed to LDA and SVM classifiers to differentiate bacterial and viral pneumonia. The performance is assessed using standard evaluation measures, including accuracy, F-measure, precision, specificity, sensitivity, and Area Under the Curve (AUC).

2.1 Image database

The chest radiographic images of 200 pneumonia (100 bacterial and 100 viral) cases are acquired from a publicly accessible dataset, “Labeled optical coherence tomography and chest X-ray images for classification” (Kermany D. et al., 2018; Kermany DS. et al., 2018). The dataset includes anteroposterior CXRs of children under 5 years. The images were acquired in the “Guangzhou Women’s and Children’s Medical Centre”. It is reported that ethical committee clearances were taken for the experimentation and data collection. The images were collected in accordance with the Declaration of Helsinki (Kermany DS. et al., 2018). For quality assurance, the images were re-viewed by experts to reduce grading errors and low-quality scans were excluded. The images are acquired in JPEG format and are subsequently resampled to 512 × 512 pixels through bicubic interpolation for standardizing the inputs for uniform analysis (Govindarajan and Swaminathan, 2019).

2.2 Hybrid model for segmentation of lungs and mediastinum

Initially, the RDLS method is employed to segment the lung regions from the CXRs (Govindarajan and Swaminathan, 2021). Equation 1 provides the growth of a Level Set Contour (LSC), ∅ t.where ∇ ∅ represents the gradient-derived normal vector, and F denotes the function that regulates the speed of the contour. During the evolution process, the LSC can become excessively steep or flattened, which can result in errors. To address this, a re-initialization step is introduced to preserve its form as a signed distance function. This procedure is often slow and unstable. To overcome the requirement of re-initialization, the RDLS approach is adopted. In this process, a diffusion term is incorporated into the level set formulation, as expressed in Equation 2.where ε denotes a constant, while Δ refers to the Laplacian operator. In this study, the parameters are set as follows: a Gaussian kernel with a standard deviation of 1.5, a fixed force of 0.7, a time increment of 0.1, and a total of 800 iterations.

∂ ∅ ∂ t + F ∇ ∅ = 0 (1)

∅ t = ε Δ ∅ + 1 ε F ∇ ∅ (2)

Additionally, the CV model is used to segment the mediastinum. Equation 3 defines the energy function corresponding to a closed contour θ within the image I x (Chung et al., 2018; Tulo et al., 2022).where c 1 and c 2 correspond to the mean intensity values inside and outside the contour, while λ 1 and λ 2 act as driving forces for the respective regions. To ensure boundary smoothness, the original energy function is altered with additional regularization terms, as described in Equation 4.where δ ∅ and H ∅ represent Dirac and Heaviside functions, respectively. Further minimization of this equation through the steepest descent approach results in a function, expressed by Equation 5.where ∇ represents the nabla operator. For accurate delineation, the process is executed for 900 iterations empirically.

E C V c 1 , c 2 , θ = λ 1 ∫ i n θ I x − c 1 2 d x + λ 2 ∫ o u t θ I x − c 2 2 d x (3)

E C V c 1 , c 2 , ∅ = λ 1 ∫ Ω I x − c 1 2 H ∅ x d x + λ 2 ∫ Ω I x − c 2 2 1 − H ∅ x d x + μ ∫ Ω δ ∅ x ∇ ∅ x d x + ν ∫ Ω H ∅ x d x (4)

∂ ∅ x , t ∂ t = δ ∅ − λ 1 I x − c 1 2 + λ 2 I x − c 2 2 + μ div ∇ ∅ ∇ ∅ − ν (5)

2.3 Biomarker extraction

Two sets of morphological biomarkers, geometric and Hu invariant moments, are extracted for characterization of structural variations in the left lung, right lung, and mediastinum due to bacterial and viral pneumonia. In total, 25 biomarkers (18 geometric and 7 Hu moments) are obtained from the delineated masks.

2.3.1 Geometric biomarkers

Eighteen geometric biomarkers are derived from each lung and mediastinum mask. The biomarker set consists of area, perimeter, orientation, equivalent diameter, circularity, eccentricity, convex area, extent, solidity, major and minor axis length, maximum and minimum Feret diameter, maximum and minimum Feret angle, bounding width and height, and bounding width height ratio. These biomarkers are extracted from the segmented regions using the methodology described in Tulo et al. (2022), which computes morphological and geometric descriptors to characterize the shape and structural properties of the regions. The descriptions of these biomarkers are provided in Table 1 (Karargyris et al., 2016; Santosh and Antani, 2018; Chandra et al., 2020).

TABLE 1. Description of the extracted biomarkers.

These biomarkers capture both local and global geometric properties of the segmented masks and collectively offer a comprehensive description of the geometry. Size-related biomarkers such as area, perimeter, and equivalent diameter quantify size and boundary length, while axis lengths, eccentricity, and orientation characterize shape and elongation. Compactness indicators, including solidity, convex area, and circularity, indicate compactness and similarity to regular forms. Extent and bounding-box dimensions capture how efficiently the object fills its enclosing frame. Feret diameters and their angles add information about directional size variation, which is helpful for irregular structures. These descriptors form a robust biomarker set for quantitative analysis and classification.

2.3.2 Hu moments

Hu moment algorithm constructs seven invariant biomarkers using central moments of orders 2 and 3 (Hu, 1962; Chevtchenko et al., 2018; Goyal and Singh, 2023). For an image I x , y, the p + q th order central moment can be given as Equation 6.where N and M represent the height and width, x ¯ and y ¯ denote centers of gravity. Further, the normalized central moment is given by Equation 7.where ρ = p + q 2 + 1. The seven invariant moments are constructed using the second and third order normalized central moments as shown in Equations 8–14 (Chevtchenko et al., 2018).

μ p q = ∑ y = 1 N ∑ x = 1 M x − x ¯ p y − y ¯ q I x , y , p , q = 0 , 1 , 2 … (6)

m p q = μ p q μ 00 ρ (7)

H 1 = m 20 + m 02 (8)

H 2 = m 20 − m 02 2 + 4 m 11 2 (9)

H 3 = m 30 − 3 m 12 2 + 3 m 21 − m 03 2 (10)

H 4 = m 30 + m 12 2 + m 21 + m 03 2 (11)

H 5 = m 30 − 3 m 12 m 30 + m 12 m 30 + m 12 2 − 3 m 21 + m 03 2 + 3 m 21 − m 03 m 21 + m 03 3 m 30 + m 12 2 − m 21 + m 03 2 (12)

H 6 = m 20 − m 02 m 30 + m 12 2 − m 21 + m 03 2 + 4 m 11 m 30 + m 12 m 21 + m 03 (13)

H 7 = 3 m 21 − m 03 m 30 + m 12 m 30 + m 12 2 − 3 m 21 + m 03 2 − m 30 − 3 m 12 m 21 + m 03 3 m 30 + m 12 2 − m 21 + m 03 2 (14)

Further, these biomarkers are assessed to identify structural variations due to bacterial and viral pneumonia. The statistical significance analysis is carried out using the Wilcoxon rank sum test. A p-value of less than 0.05 is considered statistically significant.

2.4 Biomarker fusion

The employed biomarker fusion approach is shown in Figure 2. Multiple wrapper-based biomarker selection methods are employed on the extracted left lung, right lung, and mediastinum biomarkers to obtain optimal biomarker subsets (Wang et al., 2020). These techniques evaluate subsets of biomarkers based on the performance of a chosen learning algorithm. Instead of relying on statistical scores alone, they repeatedly train and test the model with different biomarker combinations to identify the subset that produces the best predictive results. The forward biomarker selection strategy is employed, which tests performance by progressively adding biomarkers starting from an empty set. Further, it finds the set of biomarkers for which the algorithm provides the best classification result.

ML algorithms, including LDA and SVM, are used for training, and the optimal left lung, right lung, and mediastinum biomarkers are obtained based on the best performance. The dimension of the selected biomarkers depends on the employed algorithm. Further, the optimal biomarkers of the left lung, right lung, and mediastinum are fused in series connection to obtain the final set of biomarkers. In conclusion, two sets of fused biomarkers are obtained using LDA and SVM algorithms. These biomarkers are fed to the classifiers to evaluate their discrimination ability in differentiating the pneumonia conditions.

2.5 Differentiation of pneumonia conditions

The biomarkers are fed to LDA and SVM classifiers to differentiate bacterial and viral pneumonia. To enhance the distance between two classes, the LDA classifier determines the best linear transformation of the input features. It searches for a projection to maximize the inter-class to intra-class variance ratio (Rahman et al., 2019). For this classifier, the regularization parameter (gamma) value is chosen as 0. The SVM classifier generates an optimal hyperplane that maximizes the margin between feature vectors, and classification is performed according to the distance of the test sample from this hyperplane. This technique is robust and performs better even with a smaller database (Jaeger et al., 2014). In this work, a linear kernel function and a soft-margin penalty of one are used. Five-fold cross-validation is performed, and the average of these folds is reported. Classification results are analyzed through standard metrics, including accuracy, F-measure, precision, specificity, sensitivity, and AUC (Tulo et al., 2022; Rabbah et al., 2025). The metrics are calculated as given below in Equations 15–19. where True Positive (TP) refers to correctly identified viral CXRs, False Positive (FP) to bacterial images misclassified as viral, True Negative (TN) to correctly recognized bacterial images, and False Negative (FN) to viral images predicted as bacterial. The ROC curve displays the variation of the true positive rate with respect to the false positive rate, while the AUC offers a single metric of performance for all thresholds (Santosh and Antani, 2018). The work has been carried out on an AMD Ryzen 7 system with 16 GB RAM using MATLAB R2023b.

A c c u r a c y = T P + T N T P + F P + T N + F N (15)

S p e c i f i c i t y = T N T N + F P (16)

S e n s i t i v i t y = T P T P + F N (17)

P r e c i s i o n = T P T P + F P (18)

F − m e a s u r e = 2 × P r e c i s i o n × S e n s i t i v i t y P r e c i s i o n + S e n s i t i v i t y (19)

3 Results and discussion

The demonstrative bacterial and viral pneumonia CXRs are presented in Figure 3A. Visually, the left lung appears reduced in size, and the mediastinum region appears wider across the superior to inferior compartment in bacterial images than in viral images. In bacterial images, opacities are seen near the boundary between the mediastinum and right lung (shown with black arrows). This can be attributed to the occurrence of abnormalities such as consolidation (O’Grady et al., 2014; Garg et al., 2019). Meanwhile, in viral images, irregular opacities are observed in both left and right lungs (shown with white arrows), which could be due to diffuse interstitial patterns.

Figure 3B presents the final contours attained by RDLS to identify the lung borders. The approach effectively captures the lung fields with high accuracy, even in the presence of indistinct edges and structural variations across images. Figure 3C shows the segmented lung masks. The masks clearly reveal high structural variability in lungs among the pneumonia images. However, visual evaluation of these variations is complex.

Figure 4A shows the final contours attained by the CV model to identify the mediastinum borders in bacterial and viral pneumonia images. The images indicate that the approach effectively detects the mediastinum by delineating prominent landmarks, including the cardiophrenic angles and aortic knob. Figure 4B presents the binary masks of the segmented mediastinum region. The masks exhibit noticeable structural variations. The segmented masks obtained from Figures 3C, 4B are processed to extract morphological biomarkers for characterization of the alterations in both mediastinum and lungs.

The structural variations in lungs and mediastinum in terms of their area for bacterial pneumonia images are shown in Figure 5A. Both areas are found to have similar mean values. The Standard Deviation (SD) is seen to be lower for the mediastinum area than for the lungs. The Coefficient of Variation (CoV) for the mediastinum area is 32% less than the lung area. This indicates less variability of the mediastinum area around the mean as compared to lung area. This can be attributed to influence of bacterial pneumonia on lungs compared to mediastinum. The variations in viral pneumonia CXRs are illustrated in Figure 5B. Mean ± SD values of 0.45 ± 0.19 and 0.29 ± 0.12 are observed for lung area and mediastinum area, respectively. The lung area is seen to be significantly greater than mediastinum area. Both are found to have close values of CoV. This shows similar variability of the biomarkers around their mean values, indicating smaller variations in viral conditions as compared to bacterial. This also shows a high degree of enlargement of the mediastinum in bacterial pneumonia as compared to viral.

The differences between the lungs and mediastinum area for each of the bacterial and viral images are shown in Figures 5C,D, respectively. In a considered image, the greater value of mediastinum area than lung area results in a negative difference value. Approximately in 46% of bacterial pneumonia CXR images, the difference value is found to be negative, which indicates that the value of the mediastinum area is greater than the lung area. Whereas, in only 21% of images, the difference value is observed to be negative for viral conditions. This could be attributed to the mediastinal enlargement or reduction in lung area, or both, in a larger number of bacterial images than the viral images.

The comparison of variations in the lung area between bacterial and viral pneumonia images is demonstrated in Figure 6A. The mean lung area in bacterial conditions is found to be significantly smaller (p < 0.05) than in viral conditions by 14.3%. This indicates a relatively high reduction of lung area in bacterial conditions. Similar remarks are noticed for the minor axis length of lungs in Figure 6B. The biomarker in bacterial conditions is found to have a smaller mean than viral by 13.1%. Figure 6C displays the distributions of the mediastinum area. The mean area in bacterial conditions is found to be 29.6% higher than in viral conditions. This can be attributed to an enlarged mediastinum in bacterial infections. These findings are in agreement with the inferences obtained from Figure 5. From Figure 6D, the mean minor axis length of the mediastinum is observed to be 41.9% higher in bacterial images than in viral images. This indicates a comparatively widened mediastinum in bacterial conditions.

The statistical analysis results obtained using the Wilcoxon rank sum test are presented in Table 2. Biomarkers such as area, minor axis length, eccentricity, circularity, equivalent diameter, minimum Feret diameter, bounding width and height, bounding width height ratio, H 1 and H 2 are seen to be significant (p < 0.05) for both the left lung and mediastinum. The right lung orientation, extent, and maximum Feret angle are found to be significant. The significance of these biomarkers can be attributed to the distinct variations in their mean values. The mean left lung major axis length is found to be 18.3% higher in bacterial images compared to viral. A mean difference of 35.4% is observed in the bonding width of mediastinum between bacterial and viral images. Hence, it can be advocated that these biomarkers are able to significantly contribute to the characterization of pneumonia conditions.

TABLE 2. Statistical significance (p-values) of the extracted biomarkers in discriminating bacterial and viral pneumonia.

Indicates statistically significant (p* < 0.05).

Figure 7 represents the performance comparison for classifying bacterial and viral pneumonia images using mediastinum, lungs, and concatenated lungs and mediastinum biomarkers. Figure 7A shows the results obtained using the LDA classifier. Performances of greater than 76% and 77% are obtained using mediastinum and lung biomarkers, respectively. An F-measure of 80.2% is obtained using the concatenated biomarkers. Being the harmonic mean of precision and sensitivity, F-measure considers the properties of both measures. Hence, it can better evaluate the robustness and performance of the classifier. Figure 7B shows the performances obtained using the SVM classifier. F-measure of 75.5% and 81.3% are attained by employing the mediastinum and lung biomarkers, respectively. A higher F-measure of 82.9% is obtained using the concatenated lungs and mediastinum biomarkers. SVM is observed to achieve better performances using the concatenated biomarkers, which can be attributed to its capability to manage complex and large data. Overall, the concatenated lungs and mediastinum biomarkers provide higher performance than the distinct set of biomarkers.

For further improvement in the classification performances, different biomarker selection methods are employed. Table 3 shows the selected number of biomarkers using these approaches. The Wilcoxon rank sum test provided the most significant biomarkers from the left lung, followed by the mediastinum. Meanwhile, the maximum number of optimal biomarkers is obtained from the mediastinum using the SVM wrapper method. This can be attributed to significant variations in the left lung and the adjacent mediastinum due to frequent occurrences of pneumonia manifestations. This shows that the mediastinum, along with the lungs, significantly contributes to the characterization of pneumonia conditions.

TABLE 3. Number of biomarkers selected using different approaches.

Table 4 shows the classification performances using the significant lung and mediastinum biomarkers chosen by the Wilcoxon rank sum test. F-measures of 73.4%, 76.1%, and 79.4% are obtained using the mediastinum, lungs, and concatenated biomarkers, respectively. The decrease in performance can be attributed to the inability of the method to identify the optimal biomarkers that contribute to the differentiation of pneumonia conditions. Hence, further analyses are carried out using wrapper-based biomarker fusion methods.

TABLE 4. Performance (in %) for differentiation of bacterial and viral pneumonia using significant biomarkers selected through the wilcoxon rank sum test.

Table 5 shows the performances for the classification of bacterial and viral pneumonia using the optimal biomarkers selected by the wrapper-based selection methods. The wrapper selection methods are observed to perform better with their corresponding classifiers. The LDA wrapper method with the LDA classifier is found to provide an F-measure of 76.5% and 82.0% using the optimal mediastinum and lungs biomarkers, respectively. Compared to the original sets, an increase in performance is observed using the wrapper-selected biomarkers. In addition, the maximum F-measure of 87.3% is achieved using the LDA wrapper-based biomarker fusion technique along with the respective classifier.

TABLE 5. Performance (in %) for differentiation of bacterial and viral pneumonia using optimal biomarkers selected by wrapper-based methods.

The LDA biomarker fusion method is found to provide better performance than the SVM method. This could be due to the better ability of the LDA algorithm to select optimal biomarkers by efficient dimensionality reduction. Overall, it is seen that the proposed wrapper-based biomarker fusion technique can provide better classification performances than using the original biomarker sets and individual wrapper-selected biomarkers. This can be attributed to the ability of the method to select the optimal biomarkers through better generalization, comprehensive search, estimation of biomarker dependencies, and interaction with the machine learning algorithms.

Figure 8 presents the ROC curves of the classifiers for differentiation of pneumonia CXRs using wrapper-based fusion of lungs and mediastinum biomarkers. It is found that the LDA classifier achieved a maximum AUC of 88.3% using the LDA-based biomarker fusion technique. The higher value of AUC indicates a better true-classification rate of the considered pneumonia conditions. All the classifiers also obtained an AUC of greater than 87% using the wrapper-based biomarker fusion method. Even the traditional linear classifier, such as LDA, is able to provide better performances using the fusion of biomarkers from the mediastinum and lungs.

In literature, the detection of bacterial and viral pneumonia has been explored. The current work is compared with similar studies in Table 6. However, these comparisons are not conducted on a common dataset, as no prior studies have utilized the same set of images employed in this work. As a result, the comparison serves as a general reference for performance rather than a direct quantitative benchmark. The baseline study by Kermany DS. et al. (2018) analyzed pediatric chest radiographic images using a deep network named Inception. Rajaraman et al. (2018) evaluated the performance of the Inception model to differentiate the conditions in pediatric images. A study has been performed on the Chest X-ray 14 database for detecting pneumonia conditions utilizing a multi-kernel depth-wise convolution network named MDConv (Hu et al., 2020). Ibrahim et al. (2024) proposed a deep learning approach based on pretrained AlexNet model for the classification of COVID-19, viral pneumonia, bacterial pneumonia, and normal CXRs. Benmalek et al. (2025) tested different CNN models for classification of normal, bacterial and viral pneumonia, and reported that MobileNet achieved the maximum performance. However, these studies do not provide direct classification of bacterial and viral pneumonia. In other studies, by Rahman et al. (2020), Mahmud et al. (2020), multiple pre-trained deep CNN models have also been explored to classify the pneumonia images.

TABLE 6. Comparison of performance (in %) of the current work with the existing works for differentiation of bacterial pneumonia (BP) and viral pneumonia (VP) conditions.

The obtained results show that the adopted framework using wrapper biomarker fusion of mediastinum and lung biomarkers is able to achieve classification performances of greater than 85%. The study is able to obtain better accuracy and specificity than some of the deep learning approaches, such as Inception, SqueezeNet, and VGG-19. However, other studies have achieved better performance by employing customized deep learning networks. Nevertheless, studies indicate that deep learning methods need large amounts of annotated data and powerful computer systems, which are often scarce in resource-constrained environments (Govindarajan and Swaminathan, 2019). Hence, this work utilizes a less complex algorithm to extract and fuse biomarkers from both the mediastinum and lungs that can differentiate the classes using conventional machine learning techniques.

This work has been carried out on an AMD Ryzen 7 system with 16 GB RAM using MATLAB R2023b. The pipeline is evaluated on a CPU-based system without GPU acceleration, and the segmentation technique required approximately 4 s per image due to its iterative nature. Further, the feature extraction and classification stages are computationally lightweight, requiring ∼4 s for the entire dataset of 200 images accounting ∼0.02 s per image. The method does not require large annotated datasets, large-scale training or high computational resources, making it suitable for resource-constrained environments. Thus, the proposed system could be a valuable tool for primary screening of pneumonia, especially in regions with limited resources.

4 Conclusion

Differentiation of bacterial and viral pneumonia is highly important for proper treatment. However, it is challenging to classify pneumonia conditions due to the high variability. This study aims to differentiate bacterial and viral pneumonia CXRs using wrapper-based biomarker fusion of mediastinum and lung biomarkers. For this, both regions are segmented using a hybrid segmentation approach using level set techniques. Through accurate boundary edge detection, this approach facilitates better delineation of regions. The lungs and mediastinum geometric and Hu invariant moment morphological biomarkers are found to provide significant distinctions between the pneumonia classes. LDA and SVM-based wrapper fusion methods are able to combine the optimal biomarkers and have been shown to improve the classification performance.

Overall, the adopted pipeline using fusion of mediastinum and lung biomarkers is able to provide better performance than the original biomarker sets and individual wrapper-selected biomarkers. This study is able to attain better accuracy and specificity than some of the deep learning approaches reported in the literature. Future work will investigate the integration of ensemble and hybrid learning approaches, along with explainable techniques, to enhance model performance and interpretability. Additionally, evaluation on larger and more diverse datasets will be conducted to assess generalizability. Thus, the proposed work could serve as a decision-support tool, assisting radiologists in the accurate diagnosis of bacterial and viral pneumonia in pediatric patients.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Samit Kumar Ghosh, Khalifa University, United Arab Emirates

Reviewed by: Racheal Akinbo, Federal University of Technology, Nigeria

Mohamed Ramzy Ibrahim, Arab Academy for Science, Technology and Maritime Transport (AASTMT), Egypt