U-Net: Convolutional Networks for Biomedical Image Segmentation

类型	说明
论文信息	U-Net: Convolutional Networks for Biomedical Image Segmentation Olaf Ronneberger, Philipp Fischer, and Thomas Brox
会议期刊	MICCAI 2015
作者所在机构	德国弗莱堡大学
解决的问题	生物医学分割任务
模型名	U-Net
模型结构
模型特点	1.U字型结构，上采样需要用到前四个Block生成的特征图 2.注意到卷积层No-padding，所以下采样Block产生的特征图大小与上采样后的特征图大小不一致，需要裁减
数据集	ISBI 2012 挑战赛数据集（注册页面404，无法获取）
数据增强	使用3×3的随机替换的平滑变换 每个像素的替换使用双三次差值进行计算
相关参数	SGD优化算法，momentum=0.99，batchsize=1，框架caffe 初始化：标准差为$\sqrt{\frac{2}{N}}$N2​​的高斯分布初始化权重, $N$N为输入节点个数
Softmax	$p_k\left(x\right) = \frac{e^{a_{k}\left(x\right)}}{\sum_{k^{\prime} = 1}^{K} e^{a_{k}^{\prime}\left(x\right)}}$pk​(x)=∑k′=1K​eak′​(x)eak​(x)​ 1.$x\in\Omega, \Omega\subset Z^{2}$x∈Ω,Ω⊂Z2：x为每个像素 2.$a_{k}\left(x\right)$ak​(x)：每个像素点x在对应特征通道k的得分 3.$K$K为特征通道数 $p\_{k}为对于特征通道即类K的预测分类结果$p_k为对于特征通道即类K的预测分类结果
交叉熵函数	$E = \sum_{x\in\Omega}w\left(x\right)log\left(p_{\ell\left(x\right)}\left(x\right)\right)$E=x∈Ω∑​w(x)log(pℓ(x)​(x)) 1.$\ell: \Omega \rightarrow \lbrace 1,...,K \rbrace$ℓ:Ω→{1,...,K}：每个像素的真是标签 2.$\omega:\Omega\rightarrow R$ω:Ω→R：权重图
权重图	$w\left(x\right) = w_{c}\left(x\right)+w_0\cdot exp\left(-\frac{\left(d_{1}\left(x\right)+d_{2}\left(x\right)^{2}\right)}{2\sigma^{2}}\right)$w(x)=wc​(x)+w0​⋅exp⎝⎛​−2σ2(d1​(x)+d2​(x)2)​⎠⎞​ 1.$w_{c}$wc​:平衡每个类频率的权重图 2.$d_{1}:\Omega\rightarrow R$d1​:Ω→R：距离最近细胞边界的距离 3.$d_{2}:\Omega\rightarrow R$d2​:Ω→R：距离次近细胞边界的距离 4.$w_{0}=10, \sigma \approx 5$w0​=10,σ≈5
实验	—
1	电子显微镜图像的神经元分割
2	胶质母细胞瘤与海拉细胞分割实验
结论	1.实现了不同生物医学分割应用上的非常好的性能 2.仅仅需要少量标注图像以及可接受的合理训练时间。
思考	1.卷积层No-padding导致输出图像与原图不一致，故在其他应用是可以考虑加上padding=1以获得与输入相同大小图像 2.数据集全来自ISBI，需要进一步在不同机构采集的图像上实验 3.可以考虑将此模型应用于其他任务，例如分类，和目标检测等的实验

Achitecture	Specification
Model Component
BaseConvBlock	1.Convolutional Layer：channels=output_channels, kernel_size=3, padding=0, strides=1 2.ReLu Activation Layer 3.Convolutional Layer: channels=output_channels, kernel_size=3, padding=0, strides=1 4.ReLu Activation Layer(Note: no-padding)
DownSampleBlock	1.Max Pooling Layer：pool_size=2, strides=2 2.BaseConvBlock(output_channels)
UpSampleBlock	1.Deconvolutional Layer(or Upsampling Layer)：channels=output_channels, kernel_size=2, strides=2 2.BaseConvBlock(output_channels)
Model
Contracting Path
InputBlock	BaseConvBlock(output_channels)，Output: Feature Map X1
1-th DownSampleBlock	output_channels=128, Output: Feature Map X2
2-th DownSampleBlock	output_channels=256, Output: Feature Map X3
3-th DownSampleBlock	output_channels=512，Output: Feature Map X4
4-th DownSampleBlock	output_channels=1024，Output: Feature Map X5
Expanding Path
1-th UpSampleBlock	output_channels=512 A new feature map is obtained by connecting Feature Map from deconvolutional layer and Feature Map X4(need to crop) according to the channel dimension, which is used as the input of its BaseConvBlock
2-th UpSampleBlock	output_channels=256 A new feature map is obtained by connecting Feature Map from deconvolutional layer and Feature Map X3(need to crop) according to the channel dimension, which is used as the input of the BaseConvBlock
3-th UpSampleBlock	output_channels=128 A new feature map is obtained by connecting Feature Map from deconvolutional layer and Feature Map X2(need to crop) according to the channel dimension, which is used as the input of the BaseConvBlock
4-th UpSampleBlock	output_channels=64 A new feature map is obtained by connecting Feature Map from deconvolutional layer and Feature Map X1(need to crop) according to the channel dimension, which is used as the input of the BaseConvBlock
OutputBlock	1×1 Convolutional Layer：output_channels=num_classes