Attention-Guided Hybrid CNN–Transformer Architecture for Multi-Scale Skin Lesion Classification
Computer Science · 20.11.2026 · (2026)
DOI: 10.53070/bbd.2044616 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computer Science, online erschienen am 20.11.2026. Verfasst von Sibel Barin Özkan. Die Arbeit nennt 45 Literaturangaben.
Abstract (Originalsprache)
The automatic classification of skin lesions is a critical requirement in dermatology for the development of early diagnosis and clinical decision support systems. In this context, a comprehensive evaluation of the skin lesion classification problem was performed using the HAM10000 dermoscopic image dataset. The study examined 14 different transfer learning-based convolutional neural network (CNN) models, 3 different Vision Transformer (ViT) architectures, and a novel CNN+ViT hybrid model that combines the strengths of both approaches. The proposed hybrid architecture enhances the CNN-based local feature extraction process at the channel and spatial levels using the Convolutional Block Attention Module (CBAM), while effectively modeling global context information through integrated self-attention layers. This creates a multi-level attention structure focused on discriminative lesion regions. All models were trained using a five-fold cross-validation strategy and evaluated across six complementary performance metrics: accuracy, precision, recall, specificity, F-score, and Matthews correlation coefficient. Experimental results demonstrate that the proposed CNN+ViT hybrid approach delivers consistent and superior performance across different input patch sizes. Specifically, the hybrid model, achieving 99.0% accuracy, 98.8% F-score, and 97.9% MCC, provided lower variance and higher stability compared to both pure CNN and pure ViT-based architectures. These findings demonstrate that CBAM and self-attention-supported hybrid architectures offer an effective, reliable, and clinically promising solution for the automatic classification of skin lesions.
Zitieren
APA
Özkan, S. B. (2026). Attention-Guided Hybrid CNN–Transformer Architecture for Multi-Scale Skin Lesion Classification. Computer Science. https://doi.org/10.53070/bbd.2044616
Vancouver
Özkan SB. Attention-Guided Hybrid CNN–Transformer Architecture for Multi-Scale Skin Lesion Classification. Computer Science. 2026. doi:10.53070/bbd.2044616
BibTeX
@article{zkan2026,
author = {Özkan, Sibel Barin},
title = {Attention-Guided Hybrid CNN–Transformer Architecture for Multi-Scale Skin Lesion Classification},
journal = {Computer Science},
year = {2026},
number = {2026},
doi = {10.53070/bbd.2044616}
}
EUVIMED