U-Net论文精读
1 前人工作
- 卷积神经网络的流行
- 现阶段的卷积只用于分类功能,然而在生物学图像处理中,我们需要位置信息,将分类标签关联到每一个像素
这个网络的上下文使用以及定位准确性之间存在着权衡问题。较大的补丁时需要较大的Pool,这使得定位准确率下降,而较小的补丁又会使得网络的上下文联系减弱。
2 网络结构
U-Net:
Contracting path + Expensive path
[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-n3uTT5gE-1678623012325)(C:\Users\yufeixie\AppData\Roaming\Typora\typora-user-images\image-20230312172841213.png)]
U-Net:一种全卷积网络,修改了FCN的结构,采用连续的卷积进行下采样,池化算子用上采样替代,从而提高输出的分辨率,最后,将高解析特征和上采样的输出进行拼接,从而实现更精确的输出。
设定如下:
- 网络上采样部分依然有大量的通道,从而允许网络将上下文信息传递到更高层的分辨率
- 为了预测图像边界区域中的像素,通过镜像输入图像来推断缺失的上下文。这种平铺策略对于将网络应用于大图像很重要,否则分辨率会受到 GPU 内存的限制
- 数据增强对学习不变性有帮助
- 使用加权损失,使得网络在分离背景处获得更大的损失函数权重
训练策略:
- 采用随机梯度下降来训练网络
- 对于未填充的卷积,输出图像小于输入
- 使用高动量:0.09,使得先看到的训练样本决定当前的优化更新
3 创新点
Overlap-tile stategy:
为了能够预测到黄色框所示的位置,使用镜像来推断图像的上下文,这种平铺策略有助于对大图像的处理
weighted loss:
利用距离表示形态学边界信息权重,从而来平衡类别出现的频率:
w
(
x
)
=
w
c
(
x
)
+
w
0
∗
e
x
p
(
−
(
d
1
(
x
)
+
d
1
(
x
)
)
2
2
σ
2
)
w(x) = w_c (x) + w_0*exp(- \frac{(d_{1}{(x)}+d_{1}{(x)})^2}{2\sigma^2})
w(x)=wc(x)+w0∗exp(−2σ2(d1(x)+d1(x))2)
d
1
:表示最近的细胞边界,
d
2
:表示第二近的细胞边界
d_1:表示最近的细胞边界,d_2:表示第二近的细胞边界
d1:表示最近的细胞边界,d2:表示第二近的细胞边界
4 代码实现
from typing import Dict
import torch
import torch.nn as nn
import torch.nn.functional as F
# 连续卷积,设置中间输出通道接口
class DoubleConv(nn.Sequential):
def __init__(self, in_channels, out_channels, mid_channels=None):
if mid_channels is None:
mid_channels = out_channels
super(DoubleConv, self).__init__(
nn.Conv2d(in_channels, mid_channels, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(mid_channels),
nn.ReLU(inplace=True),
nn.Conv2d(mid_channels, out_channels, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True)
)
# 下采样:双线性插值或者转置卷积,如果是双线性插值,那么在进行上采样时就需要在原输入通道的基础上除以4
class Down(nn.Sequential):
def __init__(self, in_channels, out_channels):
super(Down, self).__init__(
nn.MaxPool2d(2, stride=2),
DoubleConv(in_channels, out_channels)
)
class Up(nn.Module):
def __init__(self, in_channels, out_channels, bilinear=True):
super(Up, self).__init__()
if bilinear:
self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
self.conv = DoubleConv(in_channels, out_channels, in_channels // 2)
else:
self.up = nn.ConvTranspose2d(in_channels, in_channels // 2, kernel_size=2, stride=2)
self.conv = DoubleConv(in_channels, out_channels)
def forward(self, x1: torch.Tensor, x2: torch.Tensor) -> torch.Tensor:
x1 = self.up(x1)
# [N, C, H, W]
diff_y = x2.size()[2] - x1.size()[2]
diff_x = x2.size()[3] - x1.size()[3]
# 避免由于向下取整带来的特征图大小不匹配现象
x1 = F.pad(x1, [diff_x // 2, diff_x - diff_x // 2,
diff_y // 2, diff_y - diff_y // 2])
x = torch.cat([x2, x1], dim=1)
x = self.conv(x)
return x
class OutConv(nn.Sequential):
def __init__(self, in_channels, num_classes):
super(OutConv, self).__init__(
nn.Conv2d(in_channels, num_classes, kernel_size=1)
)
class UNet(nn.Module):
def __init__(self,
in_channels: int = 1,
num_classes: int = 2,
bilinear: bool = True,
base_c: int = 64):
super(UNet, self).__init__()
self.in_channels = in_channels
self.num_classes = num_classes
self.bilinear = bilinear
self.in_conv = DoubleConv(in_channels, base_c)
self.down1 = Down(base_c, base_c * 2)
self.down2 = Down(base_c * 2, base_c * 4)
self.down3 = Down(base_c * 4, base_c * 8)
factor = 2 if bilinear else 1
self.down4 = Down(base_c * 8, base_c * 16 // factor)
self.up1 = Up(base_c * 16, base_c * 8 // factor, bilinear)
self.up2 = Up(base_c * 8, base_c * 4 // factor, bilinear)
self.up3 = Up(base_c * 4, base_c * 2 // factor, bilinear)
self.up4 = Up(base_c * 2, base_c, bilinear)
self.out_conv = OutConv(base_c, num_classes)
def forward(self, x: torch.Tensor) -> Dict[str, torch.Tensor]:
x1 = self.in_conv(x)
x2 = self.down1(x1)
x3 = self.down2(x2)
x4 = self.down3(x3)
x5 = self.down4(x4)
x = self.up1(x5, x4)
x = self.up2(x, x3)
x = self.up3(x, x2)
x = self.up4(x, x1)
logits = self.out_conv(x)
return {"out": logits}
5 总结
U-Net 架构在非常不同的生物医学分割应用中取得了良好的性能。由于数据增强算法的使用,使得很少数据的图像就可以实现较好的分割效果。