Multiclass Skin Lesions Classification and Early Skin Cancer Detection Using an Attention-Enhanced Xception Architecture
ASEAN Journal on Science and Technology for Development · 01.01.2026 · 43 (3)
DOI: 10.61931/2224-9028.1668 ↗Kurz gefasst
Fachartikel in der Zeitschrift ASEAN Journal on Science and Technology for Development, online erschienen am 01.01.2026. Verfasst von Shamira Arshad Shaikh und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Skin disorders are considered one of the major healthcare problems globally and need proper, prompt, effective and early diagnosis to enhance clinical practice. Traditional machine learning and deep learning approaches often exhibit limited generalization capability and reduced effectiveness in extracting discriminative features from complex multiclass dermatological datasets. For this purpose, an interpretable attention-enhanced Xception framework for automatic multiclass skin disease diagnosis based on dermoscopic images has been designed through the use of the International Skin Imaging Collaboration (ISIC) dataset. The dataset includes 2,357 dermoscopic images that belong to nine classes of skin lesions. The proposed framework integrates channel attention and spatial attention modules within the Xception architecture to enhance lesion localization and discriminative feature extraction. For evaluating different deep learning networks, five popular types have been used: conventional Convolutional Neural Network (CNN), VGG16, EfficientNet-B2, Baseline Xception, and Attention-Enhanced Xception architecture incorporating channel and spatial attention modules. In order to alleviate class imbalance and increase generalizability, data balancing techniques such as image rotation, horizontal flipping, zooming, and increasing brightness are used before the training process. Experimental evaluation demonstrated that the proposed attention-enhanced Xception framework achieved the highest multiclass classification accuracy of 93%, outperforming conventional CNN-based and transfer-learning architectures. In addition, precision, recall, F1-score, confusion matrix analysis, and Grad-CAM-based interpretability evaluation confirmed that the integrated attention mechanism improved discriminative feature localization and enhanced lesion-specific representation learning while suppressing irrelevant background artefacts. The obtained results indicate that the proposed framework provides a computationally efficient, interpretable, and reliable solution for automated dermatological diagnosis and computer-aided clinical decision-support systems.
Zitieren
APA
Shaikh, S. A., & Mistry, Y. (2026). Multiclass Skin Lesions Classification and Early Skin Cancer Detection Using an Attention-Enhanced Xception Architecture. ASEAN Journal on Science and Technology for Development, 43(3). https://doi.org/10.61931/2224-9028.1668
Vancouver
Shaikh SA, Mistry Y. Multiclass Skin Lesions Classification and Early Skin Cancer Detection Using an Attention-Enhanced Xception Architecture. ASEAN Journal on Science and Technology for Development. 2026;43(3). doi:10.61931/2224-9028.1668
BibTeX
@article{Shaikh2026,
author = {Shaikh, Shamira Arshad and Mistry, Yogita},
title = {Multiclass Skin Lesions Classification and Early Skin Cancer Detection Using an Attention-Enhanced Xception Architecture},
journal = {ASEAN Journal on Science and Technology for Development},
year = {2026},
volume = {43},
number = {3},
doi = {10.61931/2224-9028.1668}
}
EUVIMED