ω-net:一种用于多种医学图像的二次特征提取方法

ω-net:A Secondary Feature Extraction Method for Multiple Medical Images

针对人为无法及时从医学图像发现病理特征,最终导致病情恶化的问题,提出一种二次特征提取方法(ω-net),用来对肺、肝脏、细胞及脑质瘤进行分割。首先,将全尺寸的Unet网络作为一次特征提取路径,再将上采样路径的第三层作为起始层,用来扩展二次特征提取路径,以加强特征提取能力。其次,为了建立长期的通道依赖关系以及增强特征的位置信息,在不同的阶段引入两种新型注意力机制进行针对性优化。最后,复现了10种经典网络,在医学图像领域与基准网络Unet对比,所提网络的常用指标平均交并比、敏感性、精确度和准确率最高分别提升了0.0787、0.1287、0.1216、0.0201。经过在4种数据集上的指标数值和可视化结果比较,ω-net的多项性能指标均优于其他网络,证明了该网络的有效性与优越性。

The problem of finding pathological features artificially from medical images over time,which ultimately leads to deterioration,has increased significantly.Thus,it has become a crucial area of interest for researchers.This study introduced a secondary feature extraction method(ωnet)with the ability to segment lung,liver,nucleus,and brain tumors.First,we used the fullsize Unet as the primary feature extraction path.Similarly,we used the third layer on the upsampling path as the starting layer to expand the secondary feature extraction path to enhance the feature extraction capability.Second,we introduced two newattention mechanisms at various stages for targeted optimization to establish longterm channel dependence and enhance feature location information.Finally,the study reproduced 10 classic networks.With the application of Unet in the medical imaging field,the commonly used indicators,including mean intersection of union,sensitiveness,precision,and accuracy of the proposed network,increase by 0.0787,0.1287,0.1216,and 0.0201,respectively,compared with the benchmark network.The study evaluated the effectiveness and superiority of the introduced network and compared the index values and visualization results on four types of datasets.The results showed that the introduced network outperformed the other existing networks.