U-Net原理及代码实现

U-Net是医疗领域进行语义分割的利器，随着AIGC的爆火，U-Net已成为Diffusion Model的backbone，有必要详细记录下。

原理

U-Net包含了编码器和解码器部分：

• 编码器：通过下采样，实现了特征的层次提取。该过程类似于人类视觉系统，先关注局部细节，然后逐步构建出整体的语义信息
• 解码器：通过反卷积和跳跃连接，将编码器中相应尺寸的特征图与解码器中的特征图拼接，从而实现特征的层次恢复。该过程有助于网络在解码过程中更好地利用上下文信息，提高分割的准确性

实践

原论文的输入维度对于上下采样时的大小变换不友好，因此出现了维度裁剪的情况。这里以实现下述结构的U-Net为例：

import torch
import torch.nn as nn


class UNet(nn.Module):
    def __init__(self, in_channels, out_channels):
        super(UNet, self).__init__()

        # 下采样
        def down_conv_block(in_channels, out_channels):
            return nn.Sequential(
                nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
                nn.ReLU(inplace=True),
                nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
                nn.ReLU(inplace=True)
            )

        # 上采样
        def up_conv_block(in_channels, out_channels):
            return nn.ConvTranspose2d(in_channels, out_channels, kernel_size=2, stride=2)
        
        self.pool = nn.MaxPool2d(kernel_size=2)
        
        self.encoder1 = down_conv_block(in_channels, 32)
        self.encoder2 = down_conv_block(32, 64)
        self.encoder3 = down_conv_block(64, 128)
        self.encoder4 = down_conv_block(128, 256)

        self.bottleneck = down_conv_block(256, 512)

        self.upconv4 = up_conv_block(512, 256)
        self.decoder4 = down_conv_block(512, 256)
        self.upconv3 = up_conv_block(256, 128)
        self.decoder3 = down_conv_block(256, 128)
        self.upconv2 = up_conv_block(128, 64)
        self.decoder2 = down_conv_block(128, 64)
        self.upconv1 = up_conv_block(64, 32)
        self.decoder1 = down_conv_block(64, 32)

        self.output = nn.Conv2d(32, out_channels, kernel_size=1)

    def forward(self, x):
        # Encoding path
        # 1, 32, 256, 256
        enc1 = self.encoder1(x)
        # 1, 64, 128, 128
        enc2 = self.encoder2(self.pool(enc1))
        # 1, 128, 64, 64
        enc3 = self.encoder3(self.pool(enc2))
        # 1, 256, 32, 32
        enc4 = self.encoder4(self.pool(enc3))

        # Bottleneck
        # 1, 512, 16, 16
        bottleneck = self.bottleneck(self.pool(enc4))

        # Decoding path
        # 1, 256, 32, 32
        dec4 = self.upconv4(bottleneck)
        # 1, 512, 32, 32
        dec4 = torch.cat((dec4, enc4), dim=1)
        # 1, 256, 32, 32
        dec4 = self.decoder4(dec4)

        # 1, 128, 64, 64
        dec3 = self.upconv3(dec4)
        # 1, 256, 64, 64
        dec3 = torch.cat((dec3, enc3), dim=1)
        # 1, 128, 64, 64
        dec3 = self.decoder3(dec3)

        # 1, 64, 128, 128
        dec2 = self.upconv2(dec3)
        # 1, 128, 128, 128
        dec2 = torch.cat((dec2, enc2), dim=1)
        # 1, 64, 128, 128
        dec2 = self.decoder2(dec2)

        # 1, 32, 256, 256
        dec1 = self.upconv1(dec2)
        # 1, 64, 256, 256
        dec1 = torch.cat((dec1, enc1), dim=1)
        # 1, 32, 256, 256
        dec1 = self.decoder1(dec1)

        return self.output(dec1)


model = UNet(in_channels=3, out_channels=1)
input_tensor = torch.randn(1, 3, 256, 256)  # Batch size 1, 1 channel, 572x572 image size
output_tensor = model(input_tensor)
print(output_tensor.shape)

---

参考

• U-Net: Convolutional Networks for Biomedical Image Segmentation
• mateuszbuda_brain-segmentation-pytorch_unet