摘要
针对深层U-net网络易出现梯度消失以及存在特征重用率低的问题,本文提出一种2D DenseU-net海马体分割算法框架,融合了DenseNet和U-net两种网络结构。通过在U-net中构建当前层与前面所有层的密集连接,有效缓解了深层U-net易出现梯度消失的问题,并加强了特征传播与特征复用;DenseU-net在发挥密集连接优势的同时,保持了U-net网络中下采样和上采样的长连接结构,有助于保存浅层信息。此外,针对海马体存在比例少、体积小、边缘不清晰等特点,本文对数据样本依次进行正样本数据增强、尺寸均等剪切以及去除无效样本、边缘采样的特殊处理,有效解决了正负样本失衡问题并强化了海马体细节信息,保证获取完整的特征用于网络训练。在公开数据集ADNI(Alzheimer's Disease Neuroimaging Initiative)上的实验结果表明,本文方法能够达到92.63%的平均分割Dice精度,优于传统的海马体分割方法以及目前流行的一些基于深度学习的海马体分割模型。
Aiming at the problem that the deep U-net network is prone to gradient disappearance and low feature reuse rate, this paper proposes a 2D DenseU-net hippocampus segmentation algorithm framework, which combines DenseNet and U-net network structures. By constructing the dense connection between the current layer and all the previous layers in U-net, the problem of gradient disappearance of deep U-net is easily alleviated, and feature propagation and feature reuse are enhanced;DenseU-net maintains the advantages of dense connections while maintaining a long connection structure for downsampling and upsampling in U-net networks, helping to preserve shallow information. In addition, due to the small proportion, small size and unclear edge of the hippocampus, this paper adopts positive sample data enhancement technology, size cutting technique, deletion of invalid sample technology and edge sampling technique for the data samples, which effectively balances the positive and negative samples, and magnified the hippocampus details. These pre-processing techniques ensure that the complete features are acquired for network training. The experimental results on the public data set ADNI (Alzheimer's Disease Neuroimaging Initiative) show that the proposed method can achieve an average segmentation Dice accuracy of 92.63%, which is better than the traditional hippocampus segmentation method and some popular hippocampus segmentation models based on deep learning.
作者
时佳丽
郭立君
张荣
高琳琳
李小宝
SHI Jiali;GUO Lijun;ZHANG Rong;GAO Linlin;LI Xiaobao(Faculty of Electrical Engineering and Computer Science, Ningbo University, Ningbo 315211, Zhejiang, P.R. China)
出处
《影像科学与光化学》
CAS
2019年第4期336-348,共13页
Imaging Science and Photochemistry
基金
浙江省自然科学基金资助项目(LY17F030002)
浙江省公益技术研究计划项目(LGF18F020007)资助
作者简介
通讯作者:郭立君,E-mail:guolijun@nbu.edu.cn.
