改进Attention-UNet的多尺度肝脏CT图像分割  

Improved Attention-UNet Multi-Scale Liver CT Image Segmentation

作　　者：云飞 殷雁君 YUN Fei;YIN Yan-jun(College of Computer Science and Technology,Inner Mongolia Normal University,Hohhot 010022,China)

机构地区：[1]内蒙古师范大学计算机科学技术学院,内蒙古呼和浩特010022

出　　处：《内蒙古师范大学学报（自然科学汉文版）》2023年第2期175-180,共6页Journal of Inner Mongolia Normal University(Natural Science Edition)

基　　金：内蒙古自治区自然科学基金资助项目(2021LHMS06009)。

摘　　要：针对肝脏CT图像分割任务中U-Net自下而上的特征融合方式忽略低级特征的问题,提出基于Attention-UNet的多尺度肝脏CT图像分割方法(MFFA-Net)。在Attention-UNet结构的基础上,通过加入多尺度特征提取的方法,以减少网络特征信息的丢失。在LiTS数据集上进行了大量实验。结果表明在不同评价指标下,提出的特征信息提取和融合的方法可以有效提高分割精度。To solve the problem that the U-Net bottom-up feature fusion method ignores the low-level features in the liver CT image segmentation task, a multi-scale liver CT image segmentation method (MFFANet) based on Attention-UNet was proposed. Based on the Attention-UNet structure, a multi-scale feature extraction method is proposed to reduce the loss of network feature information. A large number of experiments have been conducted on LiTS data sets. The results show that the proposed feature information extraction and fusion method can effectively improve the segmentation accuracy under different evaluation indexes.

关 键 词：肝脏分割 U-Net 多尺度特征融合 深度学习 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]