Lesion Relational Deviation Sparse Attention With Multiscale Local–Global Learning for Automated Skin Cancer Classification
International Journal of Imaging Systems and Technology · 2026-11 · 36 (6)
DOI: 10.1002/ima.70483 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Imaging Systems and Technology, online erschienen am 10.11.2026. Verfasst von Raghuramachandran Kannamma und 3 weiteren Autor:innen. Die Arbeit nennt 23 Literaturangaben.
Abstract (Originalsprache)
Accurate skin lesion classification remains difficult because lesions with different clinical categories can show similar color, texture, and structural characteristics in dermoscopic images. This study proposes DSI‐SpTransNet, with the main contribution being a Lesion Relational Deviation Sparse Attention (LRDSA) mechanism for lesion‐aware contextual modeling. Unlike conventional attention that directly measures relationships between query and key features, LRDSA constructs a lesion‐specific reference from patch representations and describes query and key features according to their deviations from this reference. The query deviation, key deviation, and their absolute difference are used to determine relational importance, followed by top‐k selection to retain relevant lesion relationships for attention aggregation. Before LRDSA, DISubNet extracts multiscale local features describing lesion boundaries, pigmentation, and texture. The LRDSA‐based contextual representation is then fused with these retained local features for classification. Experiments on three dermoscopic datasets achieved mean accuracy, precision, recall, specificity, and F1‐score of 98.14%, 98.02%, 98.04%, 98.01%, and 98.03%, respectively. Ablation and attention comparisons further demonstrate the contribution of LRDSA to classification performance.
Zitieren
APA
Kannamma, R., Kavitha, D., Sampath, M., & Devi, M. M. Y. (2026). Lesion Relational Deviation Sparse Attention With Multiscale Local–Global Learning for Automated Skin Cancer Classification. International Journal of Imaging Systems and Technology, 36(6). https://doi.org/10.1002/ima.70483
Vancouver
Kannamma R, Kavitha D, Sampath M, Devi MMY. Lesion Relational Deviation Sparse Attention With Multiscale Local–Global Learning for Automated Skin Cancer Classification. International Journal of Imaging Systems and Technology. 2026;36(6). doi:10.1002/ima.70483
BibTeX
@article{Kannamma2026,
author = {Kannamma, Raghuramachandran and Kavitha, Dhanushkodi and Sampath, Madhusudhanan and Devi, M. M. Yamuna},
title = {Lesion Relational Deviation Sparse Attention With Multiscale Local–Global Learning for Automated Skin Cancer Classification},
journal = {International Journal of Imaging Systems and Technology},
year = {2026},
volume = {36},
number = {6},
doi = {10.1002/ima.70483}
}
EUVIMED