][]  
Research article  
Development of a Neural Network-Based Classifier for  
Child-hood Pneumonia Diagnosis Using Chest X-ray  
Images  
Desarrollo de un clasificador basado en redes neuronales  
para el diagnóstico de neumonía infantil mediante  
imágenes de ra-yos X  
Fredy Munera Romero1∗  
Adrian Gómez Consuegra 1  
1
Systems Engineering Program, Faculty of Engineering, Corporación Universitaria Rafael Núñez (Uninuñez) Cartagena, 130001,  
Colombia; fmunerar21@campusuninunez.edu.co; agomezc21@campusuninunez.edu.co  
∗
Correspondence: fmunerar21@campusuninunez.edu.co  
Citation: Munera, F., Gómez, A. Development of a Neural Network-Based Classifier for Child-hood Pneumonia Diagnosis Using Chest  
X-ray Images. OnBoard Knowledge Journal 2026, 2, 8. https://doi.org/10.70554/OBJK2026.v02n01.08  
Received: 18/03/2026, Accepted: 06/06/2026, Published: 16/06/2026  
DOI: https://doi.org/10.70554/OBJK2026.v02n01.08  
Abstract: This study presents the development of a neural network-based classifier for the diagnosis of childhood  
pneumonia using chest X-ray images. The database used contains anteroposterior radiographs from retrospective cohorts  
of pediatric patients aged one to five years from the Guangzhou Women and Children’s Medical Center. Using a routine  
developed in MATLAB, seven texture features were extracted from each image: mean, standard deviation, entropy,  
contrast, correlation, energy, and homogeneity. These features were used as inputs for a feed-forward artificial neural  
network designed to classify the images as normal or associated with bacterial pneumonia. The database consisted of  
49 normal images and 48 images with bacterial pneumonia. The best-performing neural network achieved an overall  
accuracy of 96.9%, suggesting that the proposed approach may constitute an efficient and interpretable tool to support  
computer-aided diagnosis. However, the reduced size of the dataset and the lack of external validation limit the clinical  
generalizability of the model, an aspect that should be addressed in future research.  
Keywords: Classifier; Chest X-ray images; Pediatric pneumonia; Neural networks; Computer-aided diagnosis; Artificial  
intelligence  
Resumen: Este estudio presenta el desarrollo de un clasificador basado en redes neuronales para el diagnóstico de  
neumonía infantil mediante imágenes de rayos X de tórax. La base de datos utilizada contiene radiografías anteroposte-  
riores de cohortes retrospectivas de pacientes pediátricos de uno a cinco años del Centro Médico de Mujeres y Niños  
de Guangzhou. Mediante una rutina desarrollada en MATLAB, se extrajeron siete características de textura de cada  
imagen: media, desviación estándar, entropía, contraste, correlación, energía y homogeneidad. Estas características  
fueron utilizadas como entradas de una red neuronal artificial de alimentación directa, diseñada para clasificar las  
imágenes como normales o asociadas con neumonía bacteriana. La base de datos estuvo conformada por 49 imágenes  
normales y 48 imágenes con neumonía bacteriana. La red neuronal con mejor desempeño alcanzó una exactitud global  
OnBoard Knowledge Journal 2026, 2, 8.  
https://revistasescuelanaval.com/obk/  
© 2026 by authors.  
Licensed by Escuela Naval de Cadetes "Almirante Padilla", COL.  
This article is freely accessible and distributed under the terms and conditions  
of Creative Commons Attribution (https://creativecommons.org/licenses/by/4.0/).  
OnBoard Knowledge Journal 2026, 2, 8  
2 of 12  
del 96.9%, lo que sugiere que el enfoque propuesto puede constituir una herramienta eficiente e interpretable de apoyo al  
diagnóstico asistido por computador. No obstante, el tamaño reducido del conjunto de datos y la ausencia de validación  
externa limitan la generalización clínica del modelo, aspecto que deberá abordarse en investigaciones futuras.  
Palabras clave: Clasificador; Imágenes de rayos X de tórax; Neumonía infantil; Redes neuronales; Diagnóstico asistido  
por computador; Inteligencia artificial  
1. Introduction  
Pneumonia is an acute infection of the pulmonary parenchyma that can range from localized inflamma-  
tion to fluid accumulation in the lungs. Its etiology may involve viral, bacterial, or fungal infections. It affects  
both non-hospitalized and hospitalized patients and is characterized by various clinical manifestations, in-  
cluding fever and respiratory symptoms (such as cough and expectoration), alongside observable alterations  
in chest X-ray images. Globally, this disease accounts for more than 15% of all deaths in children under five  
years of age [8].  
In the United States, pneumonia led to over 500,000 emergency department visits and more than 50,000  
deaths according to 2015 data, maintaining its position among the top ten leading causes of death worldwide  
[12].  
Despite the existence of various prevention, diagnosis, and treatment measures, half of global childhood  
pneumonia deaths occur in Africa, particularly in Nigeria (210,000 deaths), the Democratic Republic of the  
Congo (132,000), and Ethiopia (114,000). In Asia, India accounts for 410,000 deaths, followed by Pakistan  
(92,000) and Afghanistan (89,000). Consequently, for every child who loses their life to pneumonia in  
a developed country, more than 2,000 die in a developing nation. According to the Clinical University  
Hospital of Santiago de Compostela (Spain) in the journal Neumo Expertos en Prevención, it is estimated  
that 150 million children develop the disease annually, with 11 million hospitalizations occurring almost  
exclusively in developing countries, such as Mexico [2]. As part of the global effort to combat such widespread  
health challenges, technology has made significant contributions. In this context, Artificial Intelligence (AI)  
plays a leading role by combining algorithms to create systems capable of providing human-like analytical  
capabilities and assistance [3].  
Currently, Chest X-ray (CXR) imaging is widely recognized as a standard tool for pneumonia diagnosis.  
However, contemporary research focusing on data processing and predictive systems can further enhance  
diagnostic accuracy. Advanced technology pro-poses the implementation of Artificial Neural Networks  
(ANN), which are biologically inspired models composed of elements that function analogously to neurons  
and are organized in a manner similar to the human brain [1].  
The primary objective of this article is to develop and evaluate a neural network-based classifier to  
support healthcare specialists in the diagnosis of pediatric pneumonia using chest X-ray images. The  
proposed model analyzes texture features extracted from radiological images to distinguish between normal  
cases and cases associated with bacterial pneumonia. In this way, the classifier is intended as a computer-  
aided decision-support tool, particularly relevant for healthcare environments where access to specialized  
radiological interpretation may be limited.  
The article is structured as follows. Section 2 presents the main contributions of the study, emphasizing  
the development of an interpretable and computationally efficient computer-aided diagnosis approach for  
childhood pneumonia. Section 3 reviews previous studies on the use of artificial intelligence, machine  
learning, and deep learning techniques for pneumonia detection from chest X-ray images. Section 4 de-  
scribes the dataset, the feature extraction process, the neural network classifier design, and the performance  
evaluation criteria used in the study. Section 5 presents the experimental findings, including the effect of  
hidden layer size on network performance, classification accuracy, confusion matrices, and training behavior.  
Section 6 analyzes the results in relation to previous research, highlighting the advantages, limitations, and  
future research directions of the proposed approach. Finally, Section 7 summarizes the main findings and  
OnBoard Knowledge Journal 2026, 2, 8  
3 of 12  
emphasizes the potential of simple feature-based neural network models as accessible and interpretable tools  
to support pediatric pneumonia diagnosis.  
2. Contributions  
This study presents a computer-aided diagnosis (CAD) system for childhood pneumonia that offers a  
balance between high performance and computational simplicity. The main contributions of this work to the  
existing body of knowledge are threefold:  
ii.  
Unlike the current trend of applying complex and often opaque deep learning mod-els (such as convo-  
lutional neural networks) directly to raw images, this research demonstrates that a classical machine  
learning approach remains highly effective. By manually engineering a small set of seven texture  
features (mean, standard deviation, entropy, contrast, correlation, energy, and homogeneity) extracted  
via a MATLAB routine, we show that it is possible to achieve diagnostic accuracy exceeding 96.9% with  
a simple feed-forward neural network. This provides a more interpretable and less computationally  
expensive alternative for institutions with limited resources.  
ii.  
ii.  
The study provides a detailed empirical analysis of the neural network’s architecture. By systematically  
varying the number of neurons in the hidden layer (from 3 to 9) and evaluating performance metrics  
like MSE, training time, and confusion matrices, we identify an optimal configuration (4 neurons)  
that maximizes accuracy while minimizing complexity and overfitting. This granular analysis offers  
practical insights for other researchers developing similar diagnostic tools.  
This work contributes a validated, low-cost tool to support clinical decision-making. The high precision  
(96.9%) achieved on a challenging dataset of pediatric chest X-rays positions this classifier as a reliable  
"second opinion" for specialists, potentially reducing diagnostic delays and improving patient outcomes,  
particularly in primary care settings where expert radiologists may not be immediately available.  
3. Related Works  
The application of artificial intelligence (AI) for pneumonia diagnosis using chest X-ray (CXR) images  
has been an active area of research, particularly in the pediatric population where the disease burden remains  
high. Recent advances in machine learning (ML) and deep learning (DL) have led to the development of  
numerous computer-aided diagnosis (CAD) systems aimed at improving diagnostic accuracy and reducing  
the workload of radiologists [13].  
Several recent studies have demonstrated the potential of various deep learning architectures for this  
task. Salamon et al. developed a capsule neural network optimized with Bayesian optimization, achieving  
an accuracy of 95.1% on pediatric CXR images, with excellent sensitivity (98.9%) for detecting pneumonia  
cases [10]. Similarly, Khadidos et al. compared DenseNet121 and EfficientNet-B0 architectures, reporting  
that EfficientNet-B0 achieved 84.6% accuracy with the added benefit of explainability through Grad-CAM  
and LIME visualizations that highlighted clinically relevant lung regions [7]. Another study evaluated  
DenseNet-169 with transfer learning, achieving 91.6% accuracy in classifying X-ray images as normal or  
pneumonia [6].  
In a study closely related to our work, Sánchez et al. also used the Guangzhou dataset to develop  
classifiers based on both neural networks and k-nearest neighbors (K-NN). Their neural network classifier  
achieved an accuracy of 96.9%, which is identical to our results, while their K-NN approach reached 89%.  
This reinforces the robustness of the neural network approach on this specific dataset and provides a direct  
point of comparison for our feature-based methodology [11].  
Ensemble methods have also shown promising results. Harib et al. proposed a Dirichlet-evidence  
ensemble approach that incorporates uncertainty and selective prediction, achieving 100% accuracy on  
decided cases while abstaining on only 1-2% of uncertain studies, demonstrating a clinically safer deployment  
strategy [4]. A study using Maccabi Healthcare Services data in Israel validated a deep learning model on  
537 pediatric CXRs, achieving 89.4% accuracy with 92.0% precision and 99.6% specificity when compared to  
radiologist consensus [5].  
   
OnBoard Knowledge Journal 2026, 2, 8  
4 of 12  
A critical observation in the literature is the predominant reliance on a single publicly available dataset.  
According to a recent scoping review by Rickard et al. that analyzed 35 studies published between 2018 and  
2025, 31 of these studies used the Kermany dataset from the Guangzhou Women and Children’s Medical  
Center—the same dataset used in the present study [9]. This raises concerns about overfitting and limited  
generalizability to broader, realworld clinical populations. The same review reported a median accuracy  
of 92.3% for binary classification (viral vs. bacterial pneumonia) and 91.8% for multiclass classification  
(normal, viral, bacterial) across the included studies [9]. A related review by Ye and Zhou further emphasizes  
that while AI tools have been validated to quickly analyze chest images and patient data, the development  
of high-quality, multicenter datasets remains essential for improving model interpretability and clinical  
adoption [13].  
In contrast to the prevailing trend of end-to-end deep learning models that operate directly on raw  
images, the present study adopts a hybrid approach. First, seven texture features (mean, standard deviation,  
entropy, contrast, correlation, energy, and homogeneity) are extracted from CXR images using a MATLAB  
routine. These features are then fed into a simple feed-forward neural network for classification. This  
approach offers two key advantages: (1) it provides a more interpretable model by relying on radiologically  
meaning-ful features, and (2) it is computationally less expensive, making it suitable for deployment in  
resource-limited settings. While deep learning models like those in [7;10] achieve high accuracy, they often  
function as "black boxes," whereas our feature-based approach maintains a clearer link to the underlying  
radiological characteristics of pneumonia.  
4. Materials and Methods  
This section describes the methodological framework used to develop and evaluate the proposed  
pediatric pneumonia classifier. All image processing, feature extraction, and neural network implementation  
procedures were carried out in MATLAB. The methodology was structured into four main phases: dataset  
preparation, feature extraction, neural network design, and performance evaluation.  
4.1. Dataset Description  
The dataset used in this study consisted of anteroposterior chest X-ray images from retrospective cohorts  
of pediatric patients aged one to five years. The images were obtained from the publicly available pediatric  
pneumonia dataset from the Guangzhou Women and Children’s Medical Center, which has been widely  
used in research on automated pneumonia detection.  
The dataset composition is as follows:  
•
•
Normal cases: 49 images from children with no radiological evidence of pneumonia.  
Pneumonia cases: 48 images from children diagnosed with bacterial pneumonia.  
This results in a total of 97 images. The dataset was randomly partitioned for each simulation run  
into training, validation, and testing subsets. No additional data augmen-tation techniques were applied to  
preserve the original image characteristics.  
4.2. Feature Extraction  
A custom MATLAB script was developed to read the chest X-ray images and extract seven texture  
features from each image. All images were processed in grayscale format. No additional preprocessing or  
data augmentation was applied in order to preserve the original characteristics of the images and maintain  
consistency across the feature extraction process.  
The following MATLAB functions were used to extract the features:  
•
•
Mean: calculates the average pixel intensity, providing information about overall image brightness.  
Standard deviation: measures the dispersion of pixel intensities, indicating image contrast and variabil-  
ity.  
•
Entropy: quantifies the randomness or texture complexity of the image, which can differentiate between  
healthy and infected tissue patterns.  
 
OnBoard Knowledge Journal 2026, 2, 8  
5 of 12  
•
Contrast, Correlation, Energy, and Homogeneity: these four features are derived from the Gray-Level  
Co-occurrence Matrix (GLCM) and provide detailed texture analysis:  
–
Contrast: measures local intensity variations.  
–
–
–
Correlation: indicates linear dependencies between pixel intensities.  
Energy: represents texture uniformity.  
Homogeneity: measures the closeness of the distribution of elements in the GLCM to the GLCM  
diagonal.  
These seven features were chosen based on their ability to characterize lung tissue texture, which differs  
between healthy and pneumonic lungs due to the presence of infiltrates and consolidations in infected cases.  
The extracted features for all images were compiled and exported to an Excel spreadsheet for subsequent  
analysis.  
Table 1 presents the minimum and maximum values obtained for each feature across the entire dataset,  
illustrating the range of variability in the input data.  
Table 1. Extracted features from chest X-ray images: minimum and maximum values.  
Mean  
Standard deviation  
Entropy  
Contrast  
Correlation  
Energy  
Homogeneity  
58.722966  
154.142016  
25.4391131  
76.0842049  
6.55155542 105815.022  
7.84158158 1056016.32  
-0.0025452  
0.0004597  
1.7772E-07  
2.4625E-06  
0.005195399  
0.013891749  
4.3. Neural Network Classifier Design  
A feed-forward artificial neural network (ANN) with a single hidden layer was implemented using  
MATLAB’s Neural Network Toolbox. This architecture was chosen for its ability to model non-linear  
relationships between input features and output classes while maintaining computational simplicity.  
4.3.1. Network Architecture  
•
•
•
Input layer: Seven neurons corresponding to the seven extracted features (mean, standard deviation,  
entropy, contrast, correlation, energy, and homogeneity).  
Hidden layer: A single hidden layer where the number of neurons was systematically varied (from 3 to  
9) to evaluate its impact on performance.  
Output layer: One neuron with a binary output, where:  
–
–
0 represents a normal diagnosis.  
1 represents bacterial pneumonia.  
4.3.2. Training Parameters  
•
•
•
•
Learning algorithm: Scaled Conjugate Gradient Backpropagation, selected for its efficiency with small  
to medium-sized datasets and its faster convergence compared with standard gradient descent.  
Transfer functions: nonlinear transfer functions were used in the hidden layer, tangent sigmoid (tansig),  
and in the output layer, log-sigmoid (logsig), to enable nonlinear decision boundaries.  
Performance metric: Mean Squared Error (MSE), calculated as the average squared difference between  
network outputs and target values. Lower MSE values indicate better model fit.  
Data partitioning: for each simulation, the dataset was randomly divided as follows:  
–
–
–
Training set: 70% of the data.  
Validation set: 15% of the data.  
Testing set: 15% of the data.  
4.3.3. Experimental Setup  
To identify the optimal network configuration, five independent simulations were performed for each  
hidden layer size, ranging from 3 to 9 neurons. In each simulation, the network was initialized with random  
 
OnBoard Knowledge Journal 2026, 2, 8  
6 of 12  
weights and biases, trained until convergence, and evaluated using the validation and testing subsets. The  
average training time, number of epochs, and MSE were recorded for each configuration. The best-performing  
network was then further analyzed using confusion matrices to assess classification accuracy, sensitivity, and  
specificity.  
4.4. Performance Evaluation  
Model performance was assessed using multiple metrics:  
•
•
Mean Squared Error (MSE): To evaluate the goodness-of-fit during training.  
Confusion matrices: To visualize classification results, showing true positives, true negatives, false  
positives, and false negatives.  
•
•
Overall accuracy: Calculated as the proportion of correctly classified cases (both normal and pneumonia)  
over the total number of cases.  
Training time and epochs: To evaluate computational efficiency.  
The results from all configurations were compared to select the optimal network architecture, balancing  
high accuracy, low computational cost, and minimal overfitting.  
5. Results  
This section presents the experimental results obtained from the neural network classifier. Performance  
was evaluated across different network configurations using Mean Squared Error (MSE), training time,  
number of epochs, and classification accuracy as key metrics. Unless otherwise stated, results are reported as  
averages over five independent simulation runs for each configuration.  
5.1. Effect of Hidden Layer Size on Network Performance  
To determine the optimal architecture, the number of neurons in the hidden layer was varied from 3 to  
9. For each configuration, five simulations were conducted, and the average MSE, training time, and number  
of epochs were recorded.  
Table 2 summarizes the performance metrics for each hidden layer size.  
Table 2. Influence of the number of neurons in the hidden layer on network performance (average values over five  
simulations).  
Number of neurons MSE (average) Training time (s) Epochs  
3
4
5
6
7
8
9
0.36304  
0.00323  
0.28748  
0.005339  
0.38392  
0.20977  
0.22833  
3
4
5
4
4
4
2
26  
21  
8
15  
8
12  
3
As shown in Table 2, the configuration with 4 neurons in the hidden layer achieved the lowest average  
MSE (0.00323), indicating the best fit between network outputs and target values. The configuration with 6  
neurons also performed well (MSE = 0.005339), while configurations with 3, 5, 7, 8, and 9 neurons exhibited  
considerably higher error rates. Training times were consistently low across all configurations, with the  
9-neuron network training the fastest (2 seconds) but at the cost of higher MSE.  
Figure 1 illustrates the optimal network architecture identified through this experimental process: a  
feed-forward neural network with seven input features, four neurons in a single hidden layer, and one binary  
output neuron.  
   
OnBoard Knowledge Journal 2026, 2, 8  
7 of 12  
Figure 1. Optimal neural network architecture: seven input features, four hidden neurons, and one output neuron.  
Figure 2. Confusion matrix for the network with three hidden neurons.  
5.2. Classification Accuracy and Confusion Matrices  
To further evaluate the performance of the best configurations, confusion matrices were generated for  
the networks with 3, 4, and 6 hidden neurons. These matrices provide a detailed view of correct and incorrect  
classifications for each class (normal and pneumonia).  
Figure 2 presents the confusion matrix for the network with 3 hidden neurons.  
The network with 3 hidden neurons achieved an overall accuracy of 89.7%. The highest performance  
was observed in the training subset, while validation and test accuracies were slightly lower.  
Figure 3 presents the confusion matrix for the optimal network with 4 hidden neurons.  
The network with 4 hidden neurons achieved an overall accuracy of 96.9%, with the highest performance  
observed in the validation subset. This configuration correctly classified the vast majority of both normal and  
pneumonia cases, with minimal false positives and false negatives.  
Figure 4 presents the confusion matrix for the network with 6 hidden neurons.  
The network with 6 hidden neurons achieved an overall accuracy of 93.9%. Notably, this configuration  
achieved 100% accuracy on both validation and test subsets, indicating excellent generalization, although the  
training accuracy was slightly lower.  
5.3. Training Performance of the Optimal Network  
Figure 5 shows the Mean Squared Error progression during training for one representative simulation  
of the optimal 4-neuron network. The plot demonstrates the convergence behavior, with the validation error  
stabilizing at a low value, indicating successful learning without significant overfitting.  
   
OnBoard Knowledge Journal 2026, 2, 8  
8 of 12  
Figure 3. Confusion matrix for the network with four hidden neurons.  
Figure 4. Confusion matrix for the network with six hidden neurons.  
   
OnBoard Knowledge Journal 2026, 2, 8  
9 of 12  
Figure 5. Mean Squared Error during training for the four-neuron network in a representative simulation.  
6. Discussion  
This study shows that a simple feed-forward neural network trained on seven manually extracted  
texture features from chest X-ray images can achieve high classification accuracy (96.9%) in distinguishing  
normal cases from bacterial pneumonia cases in a pediatric dataset. These results are relevant not only because  
of their diagnostic performance, but also because of their implications for model simplicity, computational  
efficiency, and interpretability in potential clinical decision-support settings.  
6.1. Interpretation of Findings and Comparison with Previous Studies  
The experimental results revealed that network performance is highly sensitive to the number of neurons  
in the hidden layer. The configuration with four neurons achieved the lowest MSE (0.00323) and the highest  
overall accuracy (96.9%), providing an optimal balance between underfitting and overfitting.  
The 96.9% accuracy achieved in this study compares favorably with recent literature. (Salamon &  
˙
Ksia˛zek, 2025) reported 95.1% accuracy using capsule neural networks, while (Zhang et al., 2022) achieved  
96.91% accuracy with fine-grained convolutional neural networks, results remarkably close to ours but  
achieved with significantly more complex architectures. Other studies have reported lower accuracies:  
(Khadidos et al., 2026) ob-tained 84.6% with EfficientNet-B0, and [6] achieved 91.6% with Dense-Net-169.  
Notably, all these studies employed deep learning architectures operating directly on raw images,  
requiring substantial computational resources. In contrast, our approach achieves comparable accuracy using  
a simple neural network trained in seconds, demonstrating that thoughtful feature engineering remains  
valuable, particularly when computational resources or data are limited.  
6.2. Advantages and Limitations  
The proposed method offers three distinct advantages: first, interpretability as decisions can be traced to  
radiologically meaningful features. Second, computational efficiency with training times under five seconds  
and third, data efficiency achieving high accuracy with only 97 images.  
However, several limitations should be acknowledged. First, the dataset comprised only 97 images  
from a single institution, which limits the statistical robustness of the results and restricts the generalizability  
of the model to broader clinical populations. Although the achieved accuracy was high, the small dataset  
size increases the risk that performance may be influenced by the specific characteristics of the selected  
images. Second, the model performs only binary classification, distinguishing normal cases from bacterial  
pneumonia cases, and therefore does not address viral pneumonia, mixed infections, or other thoracic  
pathologies that may appear in real clinical practice. Third, no systematic feature selection procedure was  
   
OnBoard Knowledge Journal 2026, 2, 8  
10 of 12  
performed, and additional radiomic, morphological, or texture-based descriptors could potentially improve  
model performance. Fourth, the model lacks external validation on independent, multi-institutional datasets,  
which is essential before considering clinical deployment. Therefore, the results should be interpreted as  
preliminary evidence of technical feasibility rather than as definitive proof of clinical applicability.  
From a clinical perspective, the main challenge lies in translating the model’s performance from a  
controlled experimental dataset to heterogeneous real-world healthcare environments. Chest X-ray images  
may vary according to acquisition protocols, equipment quality, patient positioning, disease severity, and  
the presence of comorbidities. These variations may affect the stability of handcrafted texture features and,  
consequently, the classifier’s performance. For this reason, future validation should include larger and more  
diverse datasets, ideally from multiple institutions and different geographic contexts. Such validation would  
make it possible to assess the model’s robustness, reduce the risk of dataset specific bias, and determine  
whether the classifier can support decision-making in resource-limited settings without replacing expert  
clinical judgment.  
6.3. Future Research Directions  
Based on these findings, several directions for future research are proposed. First, the model should  
be externally validated using larger, multi-institutional datasets. Second, the classification task should  
be expanded to include multiclass scenarios, such as normal, bacterial pneumonia, and viral pneumonia.  
Third, additional texture descriptors, particularly Local Binary Patterns (LBP), should be explored. Fourth,  
systematic comparisons with deep learning approaches should be conducted, evaluating not only accuracy  
but also interpretability, computational cost, training requirements, and clinical usability.  
7. Conclusions  
This study developed and evaluated a neural network-based classifier for childhood pneumonia  
diagnosis using chest X-ray images. By extracting seven texture features from a dataset of 97 images, a  
feed-forward neural network with a single hidden layer was implemented and assessed. The optimal  
configuration, consisting of four neurons in the hidden layer, achieved an overall classification accuracy  
of 96.9% and an MSE of 0.00323. These findings suggest that combining classical texture features with a  
simple neural network architecture can provide an efficient and interpretable approach for computer-aided  
diagnosis.  
The proposed method offers practical advantages, particularly in terms of computational efficiency,  
interpretability, and data efficiency. These characteristics make it poten-tially useful in resource-limited  
healthcare settings where access to expert radiological interpretation may be constrained. Compared with  
more complex deep learning approaches reported in the literature, the results indicate that simpler feature-  
based models may still represent valuable tools in medical image analysis.  
Nevertheless, the findings should be interpreted with caution. The dataset was relatively small and  
originated from a single institution, which limits the generalizability of the model to broader clinical popula-  
tions. In addition, the current binary classification excludes viral pneumonia, mixed infections, and other  
thoracic conditions that may be encountered in routine clinical practice. Future work should focus on external  
validation using larger and more diverse datasets, expansion to multiclass classification, and the evaluation  
of additional texture descriptors such as Local Binary Patterns (LBP). Despite these limitations, the study  
provides preliminary evidence that an interpretable and computationally efficient neural network classifier  
can support the development of accessible diagnostic tools for underserved settings.  
Author Contributions: Fredy Munera Romero: Conceptualization, Methodology, Software, Validation, Formal analysis,  
Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization, Project  
administration. Adrian Gómez Consuegra: Conceptualization, Methodology, Software, Validation, Formal analysis,  
Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization, Project  
administration.  
 
OnBoard Knowledge Journal 2026, 2, 8  
11 of 12  
All authors contributed equally to this work. All authors have read and agreed to the published version of the  
manuscript. Please refer to the CRediT taxonomy for the definitions of the terms. Authorship should be limited to those  
who have made substantial contributions to the reported work.  
Funding: This research received no external funding.  
Institutional Review Board Statement: Not applicable, since the present study does not involvehuman personnel or  
animals.  
Informed Consent Statement: This study is limited to the use of technological resources, so nohuman personnel or  
animals are involved.  
Conflicts of Interest: Under the authorship of this research, it is declared that there is no conflict of interest with the  
present research.  
References  
1. Blasi Sanchiz, A. (2021). Classificaci’o d’imatges pel diagn‘ostic de pneum‘onia.  
2. Chaguendo Sanchez, C. A. and Avirama Buitrago, J. A. (2020). Plataforma m’ovil para el diagn’ostico autom’atico de  
la neumon’ia a partir de radiograf’ias tor’acicas soportada en deep learning.  
3. Godoy Francisco, V. L. (2020). Algoritmo de diagn’ostico preliminar de neumon’ia a partir de im’agenes radiogr’aficas  
del t’orax.  
4. Harib, W., Fouad, S., and Mahmoud, T. F. (2026). A dirichlet distribution-based trust-adaptive ensemble approach for  
pneumonia classification from chest x-ray images. In IEEE International Symposium on Biomedical Imaging (ISBI). IEEE.  
5. International Society for Pharmacoeconomics and Outcomes Research (2026). Evaluating a deep learning model for  
classifying pediatric pneumonia in israel. Retrieved March 6, 2026.  
6. Katreddi, S., Midatani, A., Roy, A. P., Velpuri, U., and Kasani, S. (2025). Pediatric pneumonia x-ray image classification:  
Predictive model development with densenet-169 transfer learning. Journal of Medical Artificial Intelligence, 8(0).  
7. Khadidos, A. O., Nanyonga, A., Khadidos, A. O., Mirza, O. M., and Yilmaz, M. T. (2026). Explainable deep learning  
for pediatric pneumonia detection in chest x-ray images.  
8. Naar P’erez, A. and Barreto Mart’inez, F. (2019). Modelo de red neuronal convolucional para el diagn’ostico de  
neumon’ia en im’agenes radiol’ogicas.  
9. Rickard, D., Kabir, M. A., and Homaira, N. (2025). Machine learning-based approaches for distinguishing viral and  
bacterial pneumonia in paediatrics: A scoping review. Computer Methods and Programs in Biomedicine, 268:108802.  
10. Salamon, S. and Ksia˛.zek, W. (2025). Capsule neural networks with bayesian optimization for pediatric pneumonia  
detection from chest x-ray images. Journal of Clinical Medicine, 14(20):7212.  
11. S’anchez, J. M. B., Villag’omez, J. L., Moreno, A. L., and Jim’enez, L. M. (2024). The development of a classifier based  
on neural networks and k-neighbors for pediatric pneumonia diagnosis through x-ray images.  
12. Tapia Vargas, G. R. (2015). Utilidad de la radiograf’ia de t’orax en el diagn’ostico de neumon’ia en pacientes  
pedi’atricos, de 2 a 5 a nos; del hospital iii yanahuara, arequipa–per’u.  
13. Ye, Y. and Zhou, W. (2026). Artificial intelligence in the diagnosis and prognosis of pediatric bacterial pneumonia:  
Current advances and challenges. Current Opinion in Pediatrics, 38(2):149.  
Authors’ Biography  
Fredy Munera Romero Systems Engineering Student.  
                         
OnBoard Knowledge Journal 2026, 2, 8  
12 of 12  
Adrian Gómez Consuegra Systems Engineering Student.  
Disclaimer/Editor’s Note: Statements, opinions, and data contained in all publications are solely those of the individual  
authors and contributors and not of the OnBoard Knowledge Journal and/or the editor(s), disclaiming any responsibility  
for any injury to persons or property resulting from any ideas, methods, instructions, or products referred to in the  
content.