摘要
针对传统UNet型肠道息肉分割模型分割精度不高的问题,提出一种基于改进CSWin Transformer的肠道息肉分割模型,分为编码器和解码器两大部分.首先,在编码阶段,利用带十字窗口的CSWin Transformer作为编码器提取肠道息肉影像的全局上下文信息.并在每层编码器的CSWin Transformer块中引入卷积块注意力(CBAM),增强模型对息肉区域和边缘信息的捕获能力.其次,在解码阶段,同样使用CSWin Transformer作为解码器,通过跳跃连接使编码器与解码器相连.最后,在编码器与解码器的中间层利用自感知注意模块(SAA)建立特征间的非局部信息交互.在开源的Kvasir-SEG、CVC-ClinicDB、EndoTect和CVC-ColonDB数据集上进行实验,所提方法分别获得0.888、0.927、0.904和0.911的Dice系数,同时获得0.876、0.902、0.831和0.860的MIoU.相比传统U型肠道息肉分割模型,Dice系数和MIoU分别提升了2.1%和2.5%.
To address the issue of limited segmentation accuracy in traditional UNet-type intestinal polyp segmentation models,a modified CSWin Transformer-based model for intestinal polyp segmentation is proposed.The model consists of two main parts:an encoder and a decoder.Firstly,during the encoding stage,a CSWin Transformer with a cross window is utilized as the encoder to extract global context information from intestinal polyp images.The CBAM is introduced into each layer of the encoder’s CSWin Transformer block to enhance the model’s ability to capture polyp area and edge information.Secondly,in the decoding stage,the CSWin Transformer is also employed as the decoder,and the encoder and decoder are connected through skip connections.Finally,in the middle layers of the encoder and decoder,a self-aware attention module(SAA)is applied to establish non-local information interaction between features.Experimental results on the open-source Kvasir-SEG,CVC-ClinicDB,EndoTect,and CVC-ColonDB datasets show that the proposed method achieves Dice coefficients of 0.888,0.927,0.904,and 0.911,respectively,along with MIoU values of 0.876,0.902,0.831,and 0.860.Compared to the traditional U-shaped intestinal polyp segmentation model,the Dice coefficient and MIoU have increased by 2.1%and 2.5%.
作者
赵宏
米珊
安定
ZHAO Hong;MI Shan;AN Ding(School of Computer Science and Artificial Intelligence,Lanzhou University of Technology,Lanzhou 730050,China;CCCC(Zhongwei)Big Data Technology Co.,Ltd.,Zhongwei 755000,China)
出处
《兰州理工大学学报》
2026年第2期91-98,共8页
Journal of Lanzhou University of Technology
基金
国家自然科学基金(62166025)。
