EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

HybridNoduleNet: Noise-Resilient Lung Nodule Classification Using CNN–ViT–DAE Architecture with Grad-CAM Interpretability

M. R. Venkatesh, Hasan Hussain Shahul Hameed

Engineering, Technology & Applied Science Research · 06.06.2026 · 16 (3) · 34920-34926

DOI: 10.48084/etasr.17755 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Engineering, Technology & Applied Science Research, online erschienen am 06.06.2026. Verfasst von M. R. Venkatesh und 1 weiteren Autor:innen. Die Arbeit nennt 18 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Lung nodule classification from Computed Tomography (CT) images is challenged by acquisition noise, anatomical variability, and the need for reliable model interpretability in clinical practice. The latest deep learning models achieve better classification results, but their performance drops when handling noisy data, and their explanation methods lose reliability because denoising and interpretability functions operate independently of their core operations. The research presents HybridNoduleNet as a complete noise-resistant system, which combines denoising and feature extraction with adaptive fusion and interpretability within a single learning framework. The framework sequentially performs noise modeling and denoising reconstruction, parallel Convolutional Neural Network (CNN) and Vision Transformer (ViT) feature extraction, adaptive cross-attention fusion, and Gradient-weighted Class Activation Mapping (Grad-CAM)-guided classification within a unified learning pipeline. The Denoising Autoencoder (DAE) learns noise-invariant representations through joint optimization with the classifier. It processes local and global features by running parallel CNN and ViT streams. The cross-attention fusion mechanism automatically adjusts these representations, and Grad-CAM produces spatially consistent explanations. The proposed framework is evaluated on the LUNA16 and LIDC-IDRI datasets under clean conditions as well as Gaussian and Poisson noise perturbations. Experimental results show that HybridNoduleNet consistently outperforms convolutional, transformer-based, and existing hybrid baselines, achieving classification accuracy above 92% under severe noise conditions. Improvements of up to 5.4% in recall and over 10% in noise robustness are observed, along with more stable and localized Grad-CAM activations, yielding Intersection over Union (IoU) gains exceeding 0.30 in noisy settings. These findings demonstrate that HybridNoduleNet provides a robust and interpretable solution for noise-aware lung nodule classification in low-dose CT imaging.

Zitieren

APA

Venkatesh, M. R., & Hameed, H. H. S. (2026). HybridNoduleNet: Noise-Resilient Lung Nodule Classification Using CNN–ViT–DAE Architecture with Grad-CAM Interpretability. Engineering, Technology & Applied Science Research, 16(3), 34920-34926. https://doi.org/10.48084/etasr.17755

Vancouver

Venkatesh MR, Hameed HHS. HybridNoduleNet: Noise-Resilient Lung Nodule Classification Using CNN–ViT–DAE Architecture with Grad-CAM Interpretability. Engineering, Technology & Applied Science Research. 2026;16(3):34920-34926. doi:10.48084/etasr.17755

BibTeX
@article{Venkatesh2026,
  author = {Venkatesh, M. R. and Hameed, Hasan Hussain Shahul},
  title = {HybridNoduleNet: Noise-Resilient Lung Nodule Classification Using CNN–ViT–DAE Architecture with Grad-CAM Interpretability},
  journal = {Engineering, Technology & Applied Science Research},
  year = {2026},
  volume = {16},
  number = {3},
  pages = {34920--34926},
  doi = {10.48084/etasr.17755}
}
Kostenlos starten →