Work

An Attention-Enhanced MobileViT Framework for Multi-Task Skin Lesion Classification and Segmentation

Year: 2026

Type: preprint

Abstract: Deep learning-based automated skin lesion diagnosis has advanced substantially, but balancing diagnostic accuracy and computational efficiency remains challenging for resource-constrained primary-care settings. Heavyweight segmentation and multi-task models deliver strong performance but carry prohibitive parameter overhead for practical deployment, while existing lightweight networks lack robust spatial feature enhancement and uncertainty constraints, limiting clinical diagnostic reliability. This work proposes an attention-augmented lightweight multi-task framework based on MobileViT-XS for joint skin lesion classification and semantic segmentation. The framework adopts MobileViT-XS as the shared backbone to extract multi-scale lesion features, integrates a convolutional block attention module (CBAM) to enhance lesion-related spatial representations under dense pixel-level supervision, and constructs a dual-branch multi-task architecture forming an asymmetric task promotion paradigm favoring classification optimization. A conservative prediction strategy coupled with risk stratification is further designed to mitigate overconfident outputs for ambiguous samples and improve diagnostic safety. Experiments on the ISIC 2019 dataset demonstrate that the proposed model achieves 84.37% accuracy, 79.46% Macro-F1, 0.935 Macro-AUC, 86.37% Dice coefficient, and 78.09% mIoU with only 2.62 M parameters. The multi-task mechanism significantly boosts classification performance compared with the single-task baseline, and yields competitive results against heavyweight architectures with far fewer parameters. Visualization validates precise lesion localization and favorable model interpretability. Overall, the proposed framework strikes a well-balanced trade-off among accuracy, computational efficiency, visual interpretability and clinical prediction safety, providing a deployable lightweight solution for primary-care skin lesion screening. Prospective multi-center clinical validation is required before real-world clinical application.

Source: Research Square

DOI: 10.21203/rs.3.rs-10583629/v1

Authors: Yanqiang Ge, Bingqian Lu, Haoyu Wang, Zixu Yang, Zhan Zhang

Institution: Anyang Normal University

Language: English

Topic: Cutaneous Melanoma Detection and Management

Subfield: Oncology

Field: Medicine

Domain: Health Sciences

SDG: Good health and well-being

Open Access status: green

Funder: Henan Provincial Science and Technology Research Project