Fusing Lung and Mediastinum Clues Helps AI Tell Bacterial From Viral Pneumonia in Children

Pneumonia remains one of the deadliest respiratory infections worldwide, and the World Health Organization estimates that around two million people die from the disease each year, with the majority of those deaths occurring in children younger than five years old. Distinguishing between bacterial and viral pneumonia is clinically critical because the two conditions demand different treatments, yet their radiographic appearances in young patients can be strikingly similar. A new study published in Frontiers in Signal Processing proposes a computer-aided approach that combines shape information from two anatomical regions of the chest—the lungs and the mediastinum—to tell these conditions apart on ordinary pediatric chest X-rays.

The research, led by Sukanta Kumar Tulo of the School of Electronics Engineering at Vellore Institute of Technology in India, together with colleagues at VIT and the Manipal Institute of Technology, takes a deliberately lightweight path. Instead of relying on massive deep neural networks, the team extracted a compact set of morphological imaging biomarkers from segmented lung and mediastinum masks and then fused the most informative ones using wrapper-based feature selection with classical machine learning classifiers. The approach is designed for settings where annotated data, computing power, and expert radiologist time are all in short supply.

The dataset consisted of 200 pediatric chest radiographs—100 bacterial and 100 viral pneumonia cases—drawn from a publicly available collection of images acquired at the Guangzhou Women’s and Children’s Medical Centre. The images, anteroposterior views of children under five years of age, had been reviewed by experts for quality assurance and were collected under ethical committee clearance in accordance with the Declaration of Helsinki. Each image was resampled to 512 by 512 pixels using bicubic interpolation to standardize the inputs for analysis.

Segmentation is the foundation of the pipeline, and the authors built a hybrid model that pairs two level set techniques, each suited to a different anatomical challenge. The lungs were delineated with the Reaction-Diffusion Level Set method, an edge-based approach that incorporates a diffusion term into the level set formulation and thereby eliminates the need for the slow and often unstable re-initialization step required by conventional level set contours. In the study, the method was configured with a Gaussian kernel of standard deviation 1.5, a fixed force of 0.7, a time increment of 0.1, and 800 iterations. The mediastinum—the central compartment of the thoracic cavity bounded laterally by the lungs—was segmented with the Chan-Vese model, a region-based technique that partitions the image according to mean pixel intensities inside and outside a closed contour, with additional regularization terms to ensure boundary smoothness. The mediastinum segmentation ran for 900 iterations, a number chosen empirically.

Once the masks were extracted, the team computed 25 morphological biomarkers per region: 18 geometric descriptors and 7 Hu invariant moments. The geometric set includes area, perimeter, orientation, eccentricity, circularity, equivalent diameter, convex area, solidity, extent, major and minor axis lengths, maximum and minimum Feret diameters and their angles, and bounding box dimensions. Size-related measures such as area and equivalent diameter quantify how large a structure is, while eccentricity and orientation capture elongation and tilt. Compactness indicators such as solidity and circularity describe how closely a region resembles a regular form, and Feret diameters add directional information that is useful for irregular structures. Hu moments, constructed from second- and third-order normalized central moments, are invariant to translation, scale, and rotation, which makes them robust descriptors of shape regardless of how the structure sits in the image.

Statistical analysis with the Wilcoxon rank sum test revealed clear morphological differences between the two pneumonia types. In bacterial cases, the mean lung area was significantly smaller than in viral cases by 14.3 percent, while the mean mediastinum area was 29.6 percent larger, and the mean minor axis length of the mediastinum was 41.9 percent higher, indicating a widened central compartment. The authors attribute the mediastinal enlargement to abnormalities such as bronchial dilatation and mediastinal lymphadenopathy, which enlarge the airways and lymph nodes, as well as pleural effusion, which can shift the mediastinum. In viral images, by contrast, irregular opacities appeared diffusely in both lungs, consistent with interstitial patterns. Roughly 46 percent of bacterial images showed a mediastinum area larger than the lung area, compared with only 21 percent of viral images. Biomarkers including area, minor axis length, eccentricity, circularity, equivalent diameter, minimum Feret diameter, and bounding dimensions were statistically significant for both the left lung and the mediastinum.

To squeeze the most discriminative power out of these measurements, the researchers applied wrapper-based biomarker selection, a strategy that repeatedly trains and tests a learning algorithm with different biomarker combinations to find the subset that yields the best predictive performance. Using forward selection starting from an empty set, they ran the procedure with two algorithms: Linear Discriminant Analysis, which seeks a linear projection that maximizes the ratio of between-class to within-class variance, and Support Vector Machine, which constructs an optimal hyperplane that maximizes the margin between classes. The optimal biomarkers from the left lung, right lung, and mediastinum were then concatenated in series to form fused biomarker sets, which were fed back into LDA and SVM classifiers with a linear kernel and a soft-margin penalty of one, evaluated with five-fold cross-validation.

The results showed a consistent pattern: fusing biomarkers from both regions outperformed using either region alone. With the SVM classifier, raw fused biomarkers achieved an F-measure of 82.9 percent, compared with 75.5 percent for mediastinum biomarkers alone and 81.3 percent for lung biomarkers alone. Wrapper selection improved things further. The LDA-based wrapper combined with the LDA classifier delivered F-measures of 76.5 percent for mediastinum biomarkers, 82.0 percent for lung biomarkers, and a peak of 87.3 percent for the fused set, with an accuracy of 87.0 percent, sensitivity of 89.0 percent, specificity of 85.0 percent, and an area under the ROC curve of 88.3 percent. The authors attribute the superiority of the LDA wrapper to its efficient dimensionality reduction, which helps it select biomarkers that generalize well.

Compared with deep learning studies on the same publicly available dataset, the biomarker fusion approach holds its own while demanding far less. The original Kermany benchmark using the Inception network reported 90.7 percent accuracy, and later deep models such as MD-Conv reached 91.0 percent, but these require large annotated datasets and powerful hardware. The new pipeline, by contrast, ran on an ordinary AMD Ryzen 7 system with 16 gigabytes of RAM in MATLAB, without GPU acceleration. Segmentation took roughly four seconds per image due to its iterative nature, while feature extraction and classification across the entire 200-image dataset required only about four seconds in total, or roughly 0.02 seconds per image. The authors note that their method achieved better accuracy and specificity than some deep learning approaches, including Inception, SqueezeNet, and VGG-19, although customized deep networks still lead in raw performance.

The practical appeal of the work lies in its suitability for resource-constrained environments, where deep learning infrastructure is often unavailable. Because the pipeline needs no large-scale training and runs on commodity hardware, the researchers suggest it could serve as a valuable tool for primary screening of pneumonia and as a decision-support system assisting radiologists in differentiating bacterial from viral infections in pediatric patients. The team plans to extend the framework with ensemble and hybrid learning approaches and explainable techniques, and to validate it on larger and more diverse datasets to assess generalizability. For now, the study offers a concrete demonstration that two anatomical regions are better than one, and that carefully chosen shape descriptors can rival heavyweight artificial intelligence at a fraction of the computational cost.

Subject of Research: Fusion of lung and mediastinum imaging biomarkers for differentiating bacterial and viral pneumonia in pediatric chest radiographs

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

Article References: Tulo, S. K., Kerketta, S. R., Rahul Manohar, O., & Martha, P. (2026). Detection of bacterial and viral pneumonia in pediatric chest radiographs using fusion of mediastinum and lung imaging biomarkers. Frontiers in Signal Processing, 6, Article 1831207. https://doi.org/10.3389/frsip.2026.1831207

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1831207

Keywords: pneumonia, pediatric chest radiographs, bacterial pneumonia, viral pneumonia, mediastinum, image segmentation, level set methods, Hu moments, biomarker fusion, machine learning, support vector machine, linear discriminant analysis