EUVIMEDScopri EUVIMED
Open access · CC BY-NC-ND 4.0

Pneu‐ConViT: A CLAHE‐Enhanced Hybrid CNN–Vision Transformer Ensemble for Pneumonia Classification

Ngoc Ha Pham, Giang Son Tran, Antoine Doucet

IET Image Processing · 2026-01 · 20 (1)

DOI: 10.1049/ipr2.70434 ↗

At a glance

Journal article in IET Image Processing, published online on 2026-07-17. Written by Ngoc Ha Pham and 2 co-authors. 56 references. The full text is freely available (open access, CC BY-NC-ND 4.0). Funded by: FPT University.

AbstractDati bibliograficiCiteRelated studies

Abstract (original language)

Pneumonia detection from chest X‐ray (CXR) images remains a challenging task due to subtle radiographic patterns and large variations in disease presentation. Recently, vision transformers (ViTs) have demonstrated strong capability in modelling global contextual relationships, while convolutional neural networks (CNNs) remain highly effective in extracting local spatial features. In this study, a hybrid framework named Pneu‐ConViT is proposed for automated pneumonia classification using chest radiographs. To enhance the visibility of subtle pulmonary abnormalities, contrast limited adaptive histogram equalisation is applied as a preprocessing step to improve local contrast in CXR images. The proposed model integrates multiple pre‐trained architectures, including ViT‐B/16, VGG16, DenseNet201, ResNet50V2, and EfficientNetB3, within a weighted late‐fusion ensemble strategy. This design enables the combination of global self‐attention representations from transformer models with hierarchical convolutional feature extraction from CNNs. Experiments conducted on the Kermany CXR dataset demonstrate that the proposed framework achieves an accuracy of 98.63%, an area under the ROC curve (AUC) of 99.90%, and an F1‐score of 99.08%, outperforming several individual backbone models and recent deep learning approaches. To further assess generalisation capability, the model was evaluated on the RSNA Pneumonia Detection Challenge dataset, where it achieved an accuracy of 84.50%, an AUC of 87.88%, and an F1‐score of 86.74%. Furthermore, model interpretability is investigated using gradient‐weighted class activation mapping and transformer attention visualisation, revealing that the proposed system focuses on clinically meaningful lung regions associated with pneumonia. These results indicate that the proposed hybrid CNN–ViT framework provides an effective and interpretable solution for computer‐aided pneumonia diagnosis from chest radiographs.

Funding

  • FPT University

Cite

APA

Pham, N. H., Tran, G. S., & Doucet, A. (2026). Pneu‐ConViT: A CLAHE‐Enhanced Hybrid CNN–Vision Transformer Ensemble for Pneumonia Classification. IET Image Processing, 20(1). https://doi.org/10.1049/ipr2.70434

Vancouver

Pham NH, Tran GS, Doucet A. Pneu‐ConViT: A CLAHE‐Enhanced Hybrid CNN–Vision Transformer Ensemble for Pneumonia Classification. IET Image Processing. 2026;20(1). doi:10.1049/ipr2.70434

BibTeX
@article{Pham2026,
  author = {Pham, Ngoc Ha and Tran, Giang Son and Doucet, Antoine},
  title = {Pneu‐ConViT: A CLAHE‐Enhanced Hybrid CNN–Vision Transformer Ensemble for Pneumonia Classification},
  journal = {IET Image Processing},
  year = {2026},
  volume = {20},
  number = {1},
  doi = {10.1049/ipr2.70434}
}
Start for free →