EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Efi-TransNet: A Hybrid EfficientNet and Multi-Head Self-Attention Framework for Explainable Pneumonia Detection from Chest X-Rays

Suneel Kumar Duvvuri, Raghavendra Goutham Mannava

Journal of Artificial Intelligence in Medical Applications · 2026

DOI: 10.65511/jaima.v2i1.1211 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Artificial Intelligence in Medical Applications, 2026. Verfasst von Suneel Kumar Duvvuri und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Pneumonia kills at every age but hits hardest in children and the elderly, where delayed diagnosis can deteriorate rapidly. Chest X-rays remain the standard first-line imaging tool, yet accurate interpretation is far more difficult than the modality’s simplicity suggests. Findings are often subtle, spatially diffuse, and easily confounded by normal anatomical structures. Convolutional neural networks have addressed the local side of this challenge reasonably well by identifying focal abnormalities, but their receptive fields do not naturally support the bilateral and long-range contextual reasoning that often distinguishes true pneumonia from normal pulmonary variation. To address this limitation, Efi-TransNet combines EfficientNetB0-based local texture extraction with a Multi-Head Self-Attention (MHSA) module capable of reasoning across distant pulmonary regions simultaneously. The framework was evaluated on a balanced chest X-ray dataset with independent training, validation, and test splits. The hybrid architecture consistently outperformed the EfficientNetB0 baseline, achieving 98% validation accuracy, 100% test accuracy, improved sensitivity, and a 43% reduction in validation loss. The substantial loss reduction suggests not only better predictive performance but also more confident and stable convergence behavior. Feature-space analysis using PCA, t-SNE, and LDA demonstrated clearer inter-class separation, while Grad-CAM visualizations confirmed lesion-focused attention over clinically relevant lung regions rather than background structures. A lightweight Gradio-based inference interface was further used to test the framework on previously unseen chest X-rays outside the standard evaluation pipeline, where performance remained stable.

Zitieren

APA

Duvvuri, S. K., & Mannava, R. G. (2026). Efi-TransNet: A Hybrid EfficientNet and Multi-Head Self-Attention Framework for Explainable Pneumonia Detection from Chest X-Rays. Journal of Artificial Intelligence in Medical Applications. https://doi.org/10.65511/jaima.v2i1.1211

Vancouver

Duvvuri SK, Mannava RG. Efi-TransNet: A Hybrid EfficientNet and Multi-Head Self-Attention Framework for Explainable Pneumonia Detection from Chest X-Rays. Journal of Artificial Intelligence in Medical Applications. 2026. doi:10.65511/jaima.v2i1.1211

BibTeX
@article{Duvvuri2026,
  author = {Duvvuri, Suneel Kumar and Mannava, Raghavendra Goutham},
  title = {Efi-TransNet: A Hybrid EfficientNet and Multi-Head Self-Attention Framework for Explainable Pneumonia Detection from Chest X-Rays},
  journal = {Journal of Artificial Intelligence in Medical Applications},
  year = {2026},
  doi = {10.65511/jaima.v2i1.1211}
}
Kostenlos starten →