A Hybrid VGG-16 and TabNet Model for Interpretable Lung Disease Detection from Chest X-rays in Resource-Constrained Environments
Journal of Computing and Social Informatics · 01.12.2025 · 5 (1) · 13-29
DOI: 10.33736/jcsi.10348.2026 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Computing and Social Informatics, online erschienen am 01.12.2025. Verfasst von Abraham Eseoghene Evwiekpaefe und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-SA 4.0).
Abstract (Originalsprache)
Accurate diagnosis of lung diseases via chest X-rays remains challenging due to subtle pathological patterns, class imbalance, and the opacity of conventional deep learning models. While convolutional neural networks excel in feature extraction, their "black-box" nature and poor interpretability hinder clinical trust, particularly in resource-constrained settings. To address these limitations, we propose a novel hybrid architecture integrating VGG-16 with TabNet, synergizing hierarchical spatial feature extraction with attention-driven interpretability. The model leverages VGG-16’s convolutional layers to capture granular details, while TabNet’s sequential attention masks dynamically prioritize discriminative features, quantifying their clinical relevance. Trained on a dataset of 2,590 chest X-rays (COPD, tuberculosis, pneumonia, and normal cases) from Nigerian hospitals, the model achieved state-of-the-art performance with 97% accuracy, surpassing ResNet-50 (95.7%) and standalone VGG-16 (94.7%). Preprocessing, including non-local means denoising and targeted augmentation, mitigates noise and class imbalance, yielding F1-scores exceeding 97% for COPD and pneumonia, with AUC values above 0.98 across all classes. The model’s interpretability is validated through attention maps highlighting disease-specific radiological markers, such as hyperinflation in COPD and consolidations in pneumonia, aligning with clinical expertise. Deployed as a real-time Android application optimized for low-end devices, the solution achieves inference in
Zitieren
APA
Evwiekpaefe, A. E., Ajakaiye, F., Musa, M. N., & Isa, M. M. (2025). A Hybrid VGG-16 and TabNet Model for Interpretable Lung Disease Detection from Chest X-rays in Resource-Constrained Environments. Journal of Computing and Social Informatics, 5(1), 13-29. https://doi.org/10.33736/jcsi.10348.2026
Vancouver
Evwiekpaefe AE, Ajakaiye F, Musa MN, Isa MM. A Hybrid VGG-16 and TabNet Model for Interpretable Lung Disease Detection from Chest X-rays in Resource-Constrained Environments. Journal of Computing and Social Informatics. 2025;5(1):13-29. doi:10.33736/jcsi.10348.2026
BibTeX
@article{Evwiekpaefe2025,
author = {Evwiekpaefe, Abraham Eseoghene and Ajakaiye, Fiyinfoluwa and Musa, Muhammad Nazeer and Isa, Muhammad Musa},
title = {A Hybrid VGG-16 and TabNet Model for Interpretable Lung Disease Detection from Chest X-rays in Resource-Constrained Environments},
journal = {Journal of Computing and Social Informatics},
year = {2025},
volume = {5},
number = {1},
pages = {13--29},
doi = {10.33736/jcsi.10348.2026}
}
EUVIMED