基于通道注意力机制和U-net的医学图像分割方法
2021-06-11王枫吕泽均
王枫 吕泽均



摘 要: 随着人工智能和医学大数据的发展,基于深度学习的医学图像分割技术因具有重要的应用价值和前景,已经成为目前的研究热点。为了增强特征图的语义信息,在U-net网络的基础上引入通道注意力机制,对U-net生成的特征逐通道进行压缩,将压缩后的特征逐通道计算权重,然后将该权重与原始特征相乘得出最终的特征。通过在两个不同器官的医学图像数据集上进行实验,Dice系数相较于原始U-net网络分别提高了2.7%和1.8%,验证了该方法的可行性和有效性。
关键词: 深度学习; U-net网络; 通道注意力机制; 医学图像分割
中图分类号:TP391 文献标识码:A 文章编号:1006-8228(2021)05-64-04
Medical image segmentation method based on channel attention mechanism and U-net
Wang Feng, Lv Zejun
(College of Computer Science (College of Software), Sichuan University, Chengdu, Sichuan 610000, China)
Abstract: With the development of artificial intelligence and medical big data, medical image segmentation technology based on deep learning, for its important application value and prospects, has become a current research hotspot. In order to enhance the semantic information of the feature map, a channel attention mechanism is introduce for U-net network to compress the features generated by U-net channel by channel, calculate the weights of the compressed features channel by channel, and then get the final features by multiplying the weights by the original features. The experiment on medical image data set of two different organs show that the Dice coefficient is increased by 2.7% and 1.8% respectively compared with the original U-net network, which verifies the feasibility and effectiveness of the method.
Key words: deep learning; U-net network; channel attention mechanism; medical image segmentation
0 引言
醫学图像分割是医学图像分析领域的关键步骤,它将医学图像中具有特殊含义的部分分割出来并提取相关特征,目的是为临床诊断提供参考数据,辅助医生做出更加准确的诊断和治疗方案[1]。由于医学图像具有不均匀、模糊、差异性等特点,对医学图像分割带来了较大的困难和挑战。因此,研究并找出更好的医学图像分割方法,提升医学图像分割技术水平,对于整个医疗领域和人工智能领域都具有重要的意义,对社会的发展和进步也具有不可替代的作用。
近年来,随着深度学习方法的迅速发展,基于深度学习的图像分割算法在医学图像分割任务中取得显著的成就。全卷积神经网络FCN是第一次将端到端的卷积网络推广到了语义分割的任务当中,随后被广泛使用和研究。……
