ViT-ResNet Fusion: An Explainable Hybrid Framework for High-Accuracy Multiclass Lung Disease Classification in Chest X-Rays
IEEE Access · 2026 · 14 · 1358-1372
DOI: 10.1109/access.2025.3649109 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Rahul Deva und 1 weiteren Autor:innen. Die Arbeit wurde bisher 4-mal zitiert (Crossref) und nennt 33 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Chest radiographs are widely used to clinically diagnose thoracic diseases such as lung opacity, pneumonia, and coronavirus disease 2019 due to their cost-effectiveness and easy access. However, accurate interpretation remains challenging because of overlapping anatomical structures, low contrast, and subtle disease manifestations. Classical deep learning models are effective but often exhibit overfitting, and weak interpretability, which restricts their clinical applicability. This paper presents an attention-driven hybrid framework that integrates contrast-limited adaptive histogram equalization-based enhancement with a dual-backbone architecture combining a vision transformer and a residual network. The vision transformer captures global contextual dependencies, while the residual network extracts local discriminative features. The concatenated representations are classified using a multi-layer perceptron and optimized end-to-end with the AdamW optimizer and a step learning rate scheduler. To improve transparency, gradient-weighted class activation mapping is used to highlight disease-relevant regions in chest radiographs. Experimental evaluation highlights that the proposed framework achieves 98.54% classification accuracy, outperforming state-of-the-art models, including EfficientNet 96.20%, DenseNet 97.50%, and a fine-tuned vision transformer 95.79%. To ensure generalizability, cross-dataset validation was conducted and trained on COVID-19 Radiography Dataset. The same model was tested on an independent COVID–Pneumonia–Normal chest X-ray dataset which achieved an accuracy of 94.07% and demonstrated good performance over heterogeneous imaging sources. These findings confirm the proposed framework’s robustness, interpretability, and suitability for real-time clinical decision support in both pandemic and routine diagnostic settings.
Zitieren
APA
Deva, R., & Dagur, A. (2026). ViT-ResNet Fusion: An Explainable Hybrid Framework for High-Accuracy Multiclass Lung Disease Classification in Chest X-Rays. IEEE Access, 14, 1358-1372. https://doi.org/10.1109/access.2025.3649109
Vancouver
Deva R, Dagur A. ViT-ResNet Fusion: An Explainable Hybrid Framework for High-Accuracy Multiclass Lung Disease Classification in Chest X-Rays. IEEE Access. 2026;14:1358-1372. doi:10.1109/access.2025.3649109
BibTeX
@article{Deva2026,
author = {Deva, Rahul and Dagur, Arvind},
title = {ViT-ResNet Fusion: An Explainable Hybrid Framework for High-Accuracy Multiclass Lung Disease Classification in Chest X-Rays},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {1358--1372},
doi = {10.1109/access.2025.3649109}
}
EUVIMED