「Medical Image Analysis」Note on 3D U-Net

QQ Group: 428014259
Tencent E-mail:403568338@qq.com
http://blog.csdn.net/dgyuanshaofeng/article/details/83904176

作者:Ozgun Cicek, Ahmed Abdulkadir, Soeren S. Lienkamp, Thomas Brox, Olaf Ronneberger
单位:略

0 摘要

在半自动情形,用户仅标注体数据中的某些片层,3D U-Net算法可学习这些稀疏标注,然后输出密集分割。属于交互分割。
在全自动情形,同上,但是属于自动分割。
采用elastic deformations进行高效数据扩充。

1 介绍

如图1所示,为从稀疏标注中学习,并输出密集分割的应用。

图 1

3D U-Net和2D U-Net类似,改动在于1、避免了bottlenecks,即 1 × 1 1 \times 1 1×1卷积(?);2、使用batchnorm加速收敛。

1.1 相关工作

2 网络结构

如图2所示,为3D U-Net网络结构。

图 2

类似于2D U-Net,3D U-Net网络结构包含分析路径和合成路径,即编码器和解码器,共有4个水平或分辨率。在分析路径,每层包含2个 3 × 3 × 3 3 \times 3 \times 3 3×3×3卷积,1个ReLU,1个步长为2的池化,显然,最后一个水平没有池化。在合成路径,每层包含1个步长为2,卷积核大小为 2 × 2 × 2 2 \times 2 \times 2 2×2×2上卷积(反卷积、转置卷积、分数步长卷积),2个 3 × 3 × 3 3 \times 3 \times 3 3×3×3卷积-ReLU组成的复合层。其中,3个水平使用了shortcut connections。最后,接输出通道为3的 1 × 1 × 1 1 \times 1 \times 1 1×1×1卷积,和加权softmax损失函数。共有参数19069955。在池化前,通道数进行doubling。在每一ReLU之前,使用BN。提及使用BN和batch size为1时,不使用global statistics,而使用current statistics。在加权softmax损失函数中,设置未标注体素的权重为0。因此,这很像体素级别的self-training技术。

3 执行细节

3.1 数据

Xenopus kidney embryos数据3套。利用Slicer3D对正交的xy,xz和yz片层进行手动标注。三轴均匀才有确定片层位置。0标签为"inside the tubule",1标签为"tubule",2标签为"background",3标签为"unlabeled"。每套数据仅标注了3个片层。

3.2 训练

数据扩充采用旋转、缩放和灰度值变换(gamma transformation?)和平滑密集deformation(B-spline interpolation)。加权方式为倒数中值频率?标签为3的体素,不贡献损失。7万次迭代,接近3天训练时间,设备为NVIDIA TianX。

4 实验

4.1 半自动分割

稀疏标注的3套数据都用于网络训练,然后对3套数据进行分割。如图3所示,为第三套数据的分割结果。显然,这很像体素级别的self-training,即学习标注体素,去分割未标注体素(网络前向过程),但是未标注体素的伪标签,并未加入训练。

[1] 3D U-Net Learning Dense Volumetric Segmentation from Sparse Annotation MICCAI 2016 [paper]

  • 0
    点赞
  • 0
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
Foreword Computational Medical Image Analysis has become a prominent field of research at the intersection of Informatics, Computational Sciences, and Medicine, supported by a vibrant community of researchers working in academics, industry, and clinical centers. During the past few years, Machine Learning methods have brought a revolution to the Computer Vision community, introducing novel efficient solutions to many image analysis problemsthat had long remained unsolved.For this revolution to enter the field of Medical Image Analysis, dedicated methods must be designed which take into account the specificity of medical images. Indeed, medical images capture the anatomy and physiology of patients through the measurements of geometrical, biophysical, and biochemical properties of their living tissues. These images are acquired with algorithms that exploit complex med- ical imaging processes whose principles must be well understood as well as those governing the complex structures and functions of the human body. The book Deep Learning for Medical Image Analysis edited by S. Kevin Zhou, Hayit Greenspan, and Dinggang Shen, top-notch researchers from both academia and industry in designing machine learning methods for medical image analysis, cov- ers state-of-the-art reviews of deep learning approaches for medical image analysis, including medical image detection/recognition, medical image segmentation, medi- cal image registration, computer aided diagnosis and disease quantification, to name some of the most important addressed problems. The book, which starts with an in- troduction to Convolutional Neural Networks for Computer Vision presents a set of novel deep learning methods applied to a variety of clinical problems and imaging modalities operating at various scales, including X-ray radiographies, Magnetic Res- onance Imaging, Computed Tomography, microscopic imaging, ultrasound imaging, etc. This impressive collection of excellent contributions will definitely se

“相关推荐”对你有帮助么?

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值