Multiscale Global Attention-Driven Deep Network for Skin Lesion Diagnosis
Journal of Engineering Science and Technology Review · 2026-02 · 19 (1) · 200-211
DOI: 10.25103/jestr.191.20 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Engineering Science and Technology Review, 2026-02. Verfasst von Yi Liu und 3 weiteren Autor:innen.
Abstract (Originalsprache)
Accurate diagnosis of skin diseases is essential for guiding appropriate treatment strategies and reducing the risk of misdiagnosis, ultimately improving patient outcomes.To enhance the feature representation for skin lesion classification and quantify the impact of artificial intelligence (AI) assistance on dermatologists' diagnostic ability, a deep learning-based dermatologic lesion detection and identification (DLDI) framework was proposed, integrating multiscale feature extraction with spatial-channel interaction modeling.Building on an improved ResNet101 (IRes-101) backbone, the proposed framework was equipped with a multiscale global attention module to reconstruct spatial and channel representations in a parallel manner.The feature discriminability was enhanced, while structural stability across varying data distributions was preserved.Based on the DLDI framework, a two-stage reader study was conducted to quantitatively compare dermatologists' diagnostic accuracy and intra-rater consistency under independent and AI-assisted settings.Results demonstrate that, the proposed model, IRes-101, achieves weighted area under the curve, mean sensitivity, and mean specificity of 0.976, 0.802, and 0.977 on a public validation set and 0.889, 0.724, and 0.964 on a local clinical validation set, respectively.With the assistance from IRes-101, dermatologists' overall diagnostic performance is strengthened.In lesion-level scoring analysis, diagnostic scores increase by 48.3% and 27.5% in the cancer and noncancer groups, respectively.This study provides empirical and methodological supports for the development of clinically applicable skin lesion diagnostic systems.
Zitieren
APA
Liu, Y., Lai, S., Zhang, C., & Das, S. K. (2026). Multiscale Global Attention-Driven Deep Network for Skin Lesion Diagnosis. Journal of Engineering Science and Technology Review, 19(1), 200-211. https://doi.org/10.25103/jestr.191.20
Vancouver
Liu Y, Lai S, Zhang C, Das SK. Multiscale Global Attention-Driven Deep Network for Skin Lesion Diagnosis. Journal of Engineering Science and Technology Review. 2026;19(1):200-211. doi:10.25103/jestr.191.20
BibTeX
@article{Liu2026,
author = {Liu, Yi and Lai, Siyu and Zhang, Chuan and Das, Sushant Kumar},
title = {Multiscale Global Attention-Driven Deep Network for Skin Lesion Diagnosis},
journal = {Journal of Engineering Science and Technology Review},
year = {2026},
volume = {19},
number = {1},
pages = {200--211},
doi = {10.25103/jestr.191.20}
}
EUVIMED