EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Dual-Stream of CNN and Transformer Feature-Based Cross-Attention Model for Multiclass Ocular Surface Squamous Neoplasia (OSSN) Classification

Iyyappan Sagadevan, Ruban Nersisson

IEEE Access · 2026 · 14 · 79107-79124

DOI: 10.1109/access.2026.3696683 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Iyyappan Sagadevan und 1 weiteren Autor:innen. Die Arbeit nennt 27 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Effective treatment planning and management of Ocular Surface Squamous Neoplasia (OSSN) depend on early, precise diagnosis from slit-lamp images. However, a major obstacle for automated systems is the diverse appearances that OSSN categories might take, along with small clinical variances. Although deep learning methods, including hybrid architectures, exhibit promise, a major obstacle to developing an accurate multi-class classification system remains the challenge of customising a single model to capture both precise local knowledge and a broad global context. The Dual-stream cross-Attentional CNN-Transformer Network (DACTNet), a novel hybrid deep learning model designed to enhance multi-class classification of OSSN, is presented in this article. DACTNet’s cross-attention module provides a more comprehensive and distinctive representation of ocular surface diseases by combining the complementary features learned by the CNN and ViT streams. A primary dataset of multi-class OSSN images, comprising normal, benign, pre-invasive, and invasive classes, is used to evaluate DACTNet. The individual performance of the lightweight CNN and ViT models is contrasted with that of the network as a whole. According to experimental findings, DACTNet outperforms comparable models across key classification metrics, including overall accuracy, precision, recall, and F1-score, underscoring its superior ability to distinguish among OSSN classes. It outperformed the best current state-of-the-art models in sensitivity, precision and F1 score, achieving an accuracy of 94%.

Zitieren

APA

Sagadevan, I., & Nersisson, R. (2026). Dual-Stream of CNN and Transformer Feature-Based Cross-Attention Model for Multiclass Ocular Surface Squamous Neoplasia (OSSN) Classification. IEEE Access, 14, 79107-79124. https://doi.org/10.1109/access.2026.3696683

Vancouver

Sagadevan I, Nersisson R. Dual-Stream of CNN and Transformer Feature-Based Cross-Attention Model for Multiclass Ocular Surface Squamous Neoplasia (OSSN) Classification. IEEE Access. 2026;14:79107-79124. doi:10.1109/access.2026.3696683

BibTeX
@article{Sagadevan2026,
  author = {Sagadevan, Iyyappan and Nersisson, Ruban},
  title = {Dual-Stream of CNN and Transformer Feature-Based Cross-Attention Model for Multiclass Ocular Surface Squamous Neoplasia (OSSN) Classification},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {79107--79124},
  doi = {10.1109/access.2026.3696683}
}
Kostenlos starten →