A Novel PneumoniaNet Framework Integrating Explainable AI for Pediatric Pneumonia Detection from Chest X-rays

Abstract

Pneumonia remains one of the leading causes of death among children worldwide, and therefore, it is necessary to use reliable and efficient tools to detect the disease early. In this study, a novel hybrid residual-dense deep learning architecture called PneumoniaNet is presented that detects pediatric pneumonia from chest X-ray images with high accuracy and interpretability. The proposed model is designed on the basis of the MobileNetV2 framework and utilizes fine-tuned deep layers with frozen shallow layers. This design enables learning high-level radiological features while preserving low-level visual information. The architecture has residual-dense connectivity to improve the propagation of gradients, feature reuse, and model generalization. The model was evaluated with a widely used benchmark pediatric chest X-ray dataset, and its results were comparatively analyzed with the performance of a number of state-of-the-art CNNs, such as DenseNet121, InceptionV3, ResNet50, EfficientNetB0, and AlexNet. PneumoniaNet yielded the best results with an accuracy of 93.11%, precision of 92.84%, recall of 93.09%, specificity of 92.43%, and F1-score of 94.29%. In addition, Gradient-weighted Class Activation Mapping visualizations showed focus on clinically relevant regions, improving interpretability and transparency. An ablation study of four PneumoniaNet variants, trained in the same manner, suggested that residual-dense connections along with selective layer freezing provide the most generalizable and effective feature representations. This work is unique because it combines residual-dense connectivity with selective layer freezing in a lightweight MobileNetV2 architecture, enabling effective feature reuse, facilitating gradient flow, and providing interpretable predictions—all within a computationally efficient framework.

Received: 29 November 2025 | Revised: 8 April 2026 | Accepted: 21 May 2026

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

The data that support the findings of this study are openly available in Kermany’s CXR Images Datasets [Kaggle] at https://www.kaggle.com/datasets/riyadhhalmosawi/kermanys-cxr-images-datasets.

Author Contribution Statement

Shahriar Siddique Arjon: Conceptualization, Methodology, Software, Investigation, Data curation, Writing – original draft, Writing – review & editing. Tamanna Yasmin: Software, Investigation, Resources, Writing – original draft, Writing – review & editing, Visualization. Ankur Kumar Mondol: Software, Formal analysis. Nakib Aman: Supervision, Project administration. Shabbir Mahmood: Validation.