获取途径

作者

Xuzhen Huang Yuliang Ma Xiajin Mei Mingxu Sun Shouqiang Jia

作者单位

⁴School of Electrical Engineering, University of Jinan, Jinan, Shandong 250024, China ³Shandong BetR Medical Technology Co., Ltd., Jinan, Shandong 250100, China ²Jinan Key Laboratory of Rehabilitation and Evaluation of Motor Dysfunction, The People’s Hospital of Huaiyin, Jinan, Shandong 250100, China ⁵Department of Radiology, Jinan People’s Hospital, Affiliated to Shandong First Medical University, Jinan, Shandong 250102, China ¹School of Automation, Hangzhou Dianzi University, Hangzhou, Zhejiang 310018, China

刊名

Biomedical Signal Processing and Control

年份

2026

卷号

Vol.112 Part C

页码

108675

ISSN

1746-8094

摘要

Recently, although Transformer performs well in skin lesion segmentation, the inadequacy of its self-attention mechanism for localized feature capture and secondary computational complexity limit real-time applications. This paper proposed FastTrans-Net, a hybrid architectural model of Transformer for the high throughput requirements of skin images. First, Fast SkinAugment is proposed for the scarcity of skin lesion image datasets, which realizes efficient data expansion by decoupling pixel-leve...更多

Recently, although Transformer performs well in skin lesion segmentation, the inadequacy of its self-attention mechanism for localized feature capture and secondary computational complexity limit real-time applications. This paper proposed FastTrans-Net, a hybrid architectural model of Transformer for the high throughput requirements of skin images. First, Fast SkinAugment is proposed for the scarcity of skin lesion image datasets, which realizes efficient data expansion by decoupling pixel-level augmentation and spatial-level augmentation. Then, Efficient Convolutional Multiplexing ShuffleNet is proposed to merge the local representation advantage of CNN with the global modeling capability of Transformer. In addition, the shift window hierarchical attention module is designed by combining shift window attention and initial structure to capture multi-scale contextual information through the hierarchical windowing mechanism, and carrier token initialization is introduced to achieve cross-window long-range dependency modeling, so that each window can be accessed to participate in both local and global representation learning. The experimental results show that Fast SkinAugment enables data diversity enhancement in ISIC2016, ISIC2017, and ISIC2018 datasets, and the FastTrans-Net segmentation effect is better improved compared with other deep learning models, with Dice metrics reaching 92.94 %, 88.72 %, and 88.97 %, respectively. The inference speed is improved by 2.1 times compared to the benchmark Transformer while maintaining competitive segmentation performance, reaching a real-time throughput of 50.22FPS.收起

文献类型

期刊