文献类型：期刊文献

中文题名：基于TransUNet的皮肤病分割网络

英文题名：Skin lesion segmentation network based on TransUNet

作者：刘一[1];杨萍[1];刘佳[1];王金华[1]

第一作者：刘一

机构：[1]北京联合大学智慧城市学院,北京100101

第一机构：北京联合大学智慧城市学院

年份：2026

卷号：43

期号：2

起止页码：181-188

中文期刊名：中国医学物理学杂志

外文期刊名：Chinese Journal of Medical Physics

基金：国家自然科学基金(62172045,62272049)。

语种：中文

中文关键词：皮肤病灶分割;图像分割;Transformer;注意力机制;深度学习

外文关键词：skin lesion segmentation;image segmentation;Transformer;attention mechanism;deep learning

摘要：皮肤病灶分割是医学影像分析的重要任务,针对其病灶区域形态多样,难以提取完整边界特征的特点,本文提出一种基于TransUNet的皮肤分割网络。主要改进以下3个方面:(1)改进十字双分支Transformer,使模型在十字窗口内计算自注意力,高效建模全局依赖;(2)采用多维度特征提取模块,通过多尺度大卷积核与多维注意力机制,捕捉皮肤病灶的边界信息;(3)引入深浅层融合模块,采用动态权重策略,实现特征完整融合,增强模型鲁棒性。实验表明在ISIC2017和ISIC2018数据集中,本文算法均取得最优性能。其中ISIC2017数据集Jaccard系数达到85.8%,相比基线TransUNet提高3.6%。在ISIC2018数据集中Dice相似系数达到90.3%,相比基线TransUNet提高2.8%。
Skin lesion segmentation is a crucial task in medical image analysis.Given the diverse morphologies of lesion areas and the difficulty in extracting complete boundary features,a skin segmentation network based on TransUNet is proposed.The main improvements are as follows:(1)improving the cross-shaped dual-branch Transformer to enable the model to compute self-attention within a cross-shaped window,thereby efficiently modeling global dependencies;(2)employing a mix structure module to capture boundary information of skin lesions through multi-scale large convolution kernels and multi-dimensional attention mechanisms;(3)introducing a deep-shallow layer fusion module that utilizes a dynamic weighting strategy to achieve complete feature integration and enhance model robustness.Experiments show that the proposed algorithm achieves the optimal performance on both the ISIC2017 and ISIC2018 datasets.Specifically,it attains a Jaccard coefficient of 85.8%on the ISIC2017 dataset reaches,demonstrating a 3.6%improvement over the baseline TransUNet,and yields a Dice coefficient of 90.3%on the ISIC2018 dataset,showing a 2.8%improvement as compared with the baseline TransUNet.