跨尺度跨维度的自适应Transformer网络应用于结直肠息肉分割  

Cross-scale and cross-dimensional adaptive transformer network for colorectal polyp segmentation

作　　者：梁礼明[1] 何安军 李仁杰 吴健[1] LIANG Liming;HE Anjun;LI Renjie;WU Jian(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China)

机构地区：[1]江西理工大学电气工程及其自动化学院,江西赣州341000

出　　处：《光学精密工程》2023年第18期2700-2712,共13页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.51365017,No.61463018);江西省自然科学基金面上项目资助(No.20192BAB205084);江西省教育厅科学技术研究重点项目资助(No.GJJ170491,No.GJJ2200848)。

摘　　要：针对结直肠息肉图像病灶区域尺度变化大、边界模糊、形状不规则且与正常组织对比度低等问题,导致边缘细节信息丢失和病灶区域误分割,提出一种跨尺度跨维度的自适应Transformer分割网络。该网络一是利用Transformer编码器建模输入图像的全局上下文信息,多尺度分析结直肠息肉病灶区域。二是通过通道注意力桥和空间注意力桥减少通道维度冗余和增强模型空间感知能力,抑制背景噪声。三是采用多尺度密集并行解码模块来填补各层跨尺度特征信息之间的语义空白,有效聚合多尺度上下文特征。四是设计面向边缘细节的多尺度预测模块,以可学习的方式引导网络去纠正边界错误预测分类。在CVC-ClinicDB、Kvasir-SEG、CVC-ColonDB和ETIS数据集上进行实验,其Dice相似性系数分别为0.942,0.932,0.811和0.805,平均交并比分别为0.896,0.883,0.731和0.729,其分割性能优于现有方法。仿真实验表明,本文方法能有效改善结直肠息肉病灶区域误分割,具有较高的分割精度,为结直肠息肉诊断提供新窗口。To address the problem of large-scale variation,blurred boundaries,irregular shapes,and low contrast with normal tissues in colon polyp images,which leads to the loss of edge detail information and mis-segmentation of lesion areas,we propose a cross-dimensional and cross-scale adaptive transformer seg⁃mentation network.First,the network uses transformer encoders to model the global contextual informa⁃tion of the input image and analyze the colon polyp lesion areas at multiple scales.Second,the channel at⁃tention and spatial attention bridges are used to reduce channel dimension redundancy and enhance the mod⁃el's spatial perception ability while suppressing background noise.Third,the multi-scale dense parallel de⁃coding module is used to bridge the semantic gaps between cross-scale feature information at different lay⁃ers,effectively aggregating multi-scale contextual features.Fourth,a multi-scale prediction module is de⁃signed for edge details,guiding the network to correct boundary errors in a learnable manner.The experi⁃mental results conducted on the CVC-ClinicDB,Kvasir-SEG,CVC-ColonDB,and ETIS datasets showed that the Dice similarity coefficients are 0.942,0.932,0.811,and 0.805,and the average inter⁃section-over-union ratios are 0.896,0.883,0.731,and 0.729,respectively.The segmentation perfor⁃mance of our proposed method was better than that of existing methods.The simulation experiment showed that our method can effectively improve the mis-segmentation of colon polyp lesion areas and achieve high segmentation accuracy,providing a new approach for colon polyp diagnosis.

关 键 词：结直肠息肉 TRANSFORMER 多尺度密集并行解码模块 多尺度预测模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]