摘要
针对现有肾脏及肿瘤图像语义分割算法存在边界处理不精细、小尺寸肾脏和肿瘤图像难以正确分割的问题,提出一种基于改进U-Net的肾脏及肿瘤图像语义分割算法。该算法应用通道注意力机制与空间注意力机制为目标特征赋予更高的权重,减弱背景信息的影响,并采用密集连接的方式增强加权后特征信息的重复利用。引入深监督机制,构建一种混合损失函数,增强网络对目标特征的提取能力并使浅层网络能够学习到更多的语义信息。在KITS19数据集上的实验结果表明,相比其他改进U-Net和PSPNet等语义分割算法,该算法对边界的处理更加精细,并解决了小尺寸肾脏和肿瘤图像可能出现的错误分割问题,有效提升了肾脏及肿瘤图像的分割精度。
Aimed at the problems that the existing semantic segmentation algorithm of kidney and tumor images has poor boundary processing and it is difficult to segment small-sized kidney and tumor images correctly,a semantic segmentation algorithm of kidney and tumor images based on improved U-Net is proposed.The algorithm applied channel attention mechanism and spatial attention mechanism to give higher weight to target features,weaken the influence of background information.And it used dense connection to enhance the reuse of weighted feature information.A deep supervision mechanism was introduced to construct a mixed loss function to enhance the network’s ability to extract target features and enable the shallow network to learn more semantic information.The experimental results on the KITS19 data set show that compared with other semantic segmentation algorithms,such as PSPNet and other improved algorithms based on U-Net,this algorithm handles the boundary more finely,solves the problem of incorrect segmentation of small-sized kidney and tumor images and effectively improves the accuracy of segmentation of kidney and tumor images.
作者
柳阔
田景文
Liu Kuo;Tian Jingwen(School of Smart City,Beijing Union University,Beijing 100101,China)
出处
《计算机应用与软件》
北大核心
2022年第9期240-247,共8页
Computer Applications and Software
基金
国家自然科学基金项目(51404074)。
关键词
语义分割
注意力机制
深监督
Semantic segmentation
Attention mechanism
Deep supervision
