【醫學圖像分割網絡】之U-Net網絡PyTorch複現

1.内容

U-Net網絡算是醫學圖像分割領域的開山之作，我接觸深度學習到現在大概将近大半年時間，看到了很多基于U-Net網絡的變體，後續也會繼續和大家一起分享學習。我現在正常拿到一個分割任務時，都是先拿U-Net網絡進行測試，一般U-Net跑出的模型精度是非常高的，雖說論文出的早，但是确實經典，在很多分割任務上表現不俗，優點和網絡結構就不談了，很多部落格都有詳細的介紹，論文也說的很清楚，這裡直接上代碼。

這裡提出的是，很多部落格都不是按照原始作者論文裡的網絡結構，比如在卷積的padding，跳層連接配接的尺寸上都做了一定的修改，雖然這樣更适合多種資料的訓練，更加使用，但是對于我們學會複現原始結構造成了障礙，這裡我分享出兩種，一種是原始論文的，一種是更加通用的。

2.論文原始版U-Net實作方式

"""
論文原始版U-Net
"""
import torch
from torch import nn
import torch.nn.functional as F
from torchsummary import summary
from collections import OrderedDict

# 編碼連續卷積層
def contracting_block(in_channels, out_channels):
    block = torch.nn.Sequential(
        nn.Conv2d(kernel_size=(3, 3), in_channels=in_channels, out_channels=out_channels),
        nn.ReLU(),
        nn.BatchNorm2d(out_channels),
        nn.Conv2d(kernel_size=(3, 3), in_channels=out_channels, out_channels=out_channels),
        nn.ReLU(),
        nn.BatchNorm2d(out_channels)
    )
    return block

# 解碼上采樣卷積層
class expansive_block(nn.Module):
    def __init__(self, in_channels, mid_channels, out_channels):
        super(expansive_block, self).__init__()

        self.up = nn.ConvTranspose2d(in_channels, in_channels // 2, kernel_size=(3, 3), stride=2, padding=1,
                                     output_padding=1, dilation=1)

        self.block = nn.Sequential(
            nn.Conv2d(kernel_size=(3, 3), in_channels=in_channels, out_channels=mid_channels),
            nn.ReLU(),
            nn.BatchNorm2d(mid_channels),
            nn.Conv2d(kernel_size=(3, 3), in_channels=mid_channels, out_channels=out_channels),
            nn.ReLU(),
            nn.BatchNorm2d(out_channels)
        )

    def forward(self, e, d):
        d = self.up(d)
        # concat
        diffY = e.size()[2] - d.size()[2]
        diffX = e.size()[3] - d.size()[3]
        e = e[:, :, diffY // 2:e.size()[2] - diffY // 2, diffX // 2:e.size()[3] - diffX // 2]
        cat = torch.cat([e, d], dim=1)
        out = self.block(cat)
        return out

# 輸出層
def final_block(in_channels, out_channels):
    block = nn.Sequential(
        nn.Conv2d(kernel_size=(1, 1), in_channels=in_channels, out_channels=out_channels),
        nn.ReLU(),
        nn.BatchNorm2d(out_channels),
    )
    return block

# U-Net
class Unet(nn.Module):

    def __init__(self, in_channel, out_channel):
        super(Unet, self).__init__()
        # Encode
        self.conv_encode1 = contracting_block(in_channels=in_channel, out_channels=64)
        self.conv_pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv_encode2 = contracting_block(in_channels=64, out_channels=128)
        self.conv_pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv_encode3 = contracting_block(in_channels=128, out_channels=256)
        self.conv_pool3 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv_encode4 = contracting_block(in_channels=256, out_channels=512)
        self.conv_pool4 = nn.MaxPool2d(kernel_size=2, stride=2)
        # Bottleneck
        self.bottleneck = torch.nn.Sequential(
            nn.Conv2d(kernel_size=3, in_channels=512, out_channels=1024),
            nn.ReLU(),
            nn.BatchNorm2d(1024),
            nn.Conv2d(kernel_size=3, in_channels=1024, out_channels=1024),
            nn.ReLU(),
            nn.BatchNorm2d(1024)
        )
        # Decode
        self.conv_decode4 = expansive_block(1024, 512, 512)
        self.conv_decode3 = expansive_block(512, 256, 256)
        self.conv_decode2 = expansive_block(256, 128, 128)
        self.conv_decode1 = expansive_block(128, 64, 64)
        self.final_layer = final_block(64, out_channel)

    def forward(self, x):
        # set_trace()
        # Encode
        encode_block1 = self.conv_encode1(x)
        encode_pool1 = self.conv_pool1(encode_block1)
        encode_block2 = self.conv_encode2(encode_pool1)
        encode_pool2 = self.conv_pool2(encode_block2)
        encode_block3 = self.conv_encode3(encode_pool2)
        encode_pool3 = self.conv_pool3(encode_block3)
        encode_block4 = self.conv_encode4(encode_pool3)
        encode_pool4 = self.conv_pool4(encode_block4)

        # Bottleneck
        bottleneck = self.bottleneck(encode_pool4)

        # Decode
        decode_block4 = self.conv_decode4(encode_block4, bottleneck)
        decode_block3 = self.conv_decode3(encode_block3, decode_block4)
        decode_block2 = self.conv_decode2(encode_block2, decode_block3)
        decode_block1 = self.conv_decode1(encode_block1, decode_block2)

        final_layer = self.final_layer(decode_block1)

        return final_layer
           

3.實際通用版U-Net實作方式1

"""
實用版U-Net
"""


class DoubleConv(nn.Module):
    def __init__(self, in_ch, out_ch):
        super(DoubleConv, self).__init__()
        self.conv = nn.Sequential(
            nn.Conv2d(in_ch, out_ch, 3, padding=1),
            nn.BatchNorm2d(out_ch),
            nn.ReLU(inplace=True),
            nn.Conv2d(out_ch, out_ch, 3, padding=1),
            nn.BatchNorm2d(out_ch),
            nn.ReLU(inplace=True)
        )

    def forward(self, input):
        return self.conv(input)


class Unet2(nn.Module):
    def __init__(self, in_ch, out_ch):
        super(Unet2, self).__init__()

        self.conv1 = DoubleConv(in_ch, 32)
        self.pool1 = nn.MaxPool2d(2)
        self.conv2 = DoubleConv(32, 64)
        self.pool2 = nn.MaxPool2d(2)
        self.conv3 = DoubleConv(64, 128)
        self.pool3 = nn.MaxPool2d(2)
        self.conv4 = DoubleConv(128, 256)
        self.pool4 = nn.MaxPool2d(2)
        self.conv5 = DoubleConv(256, 512)
        self.up6 = nn.ConvTranspose2d(512, 256, 2, stride=2)
        self.conv6 = DoubleConv(512, 256)
        self.up7 = nn.ConvTranspose2d(256, 128, 2, stride=2)
        self.conv7 = DoubleConv(256, 128)
        self.up8 = nn.ConvTranspose2d(128, 64, 2, stride=2)
        self.conv8 = DoubleConv(128, 64)
        self.up9 = nn.ConvTranspose2d(64, 32, 2, stride=2)
        self.conv9 = DoubleConv(64, 32)
        self.conv10 = nn.Conv2d(32, out_ch, 1)

    def forward(self, x):
        c1 = self.conv1(x)
        p1 = self.pool1(c1)
        c2 = self.conv2(p1)
        p2 = self.pool2(c2)
        c3 = self.conv3(p2)
        p3 = self.pool3(c3)
        c4 = self.conv4(p3)
        p4 = self.pool4(c4)
        c5 = self.conv5(p4)

        up_6 = self.up6(c5)
        merge6 = torch.cat([up_6, c4], dim=1)
        c6 = self.conv6(merge6)
        up_7 = self.up7(c6)
        merge7 = torch.cat([up_7, c3], dim=1)
        c7 = self.conv7(merge7)
        up_8 = self.up8(c7)
        merge8 = torch.cat([up_8, c2], dim=1)
        c8 = self.conv8(merge8)
        up_9 = self.up9(c8)
        merge9 = torch.cat([up_9, c1], dim=1)
        c9 = self.conv9(merge9)
        c10 = self.conv10(c9)
        out = nn.Sigmoid()(c10)

        return out
           

4.實際通用版U-Net實作方式2

"""
實用版U-Net2
"""


class Unet3(nn.Module):

    def __init__(self, in_channels=3, out_channels=1, init_features=32):
        super(Unet3, self).__init__()

        features = init_features
        self.encoder1 = Unet3._block(in_channels, features, name="enc1")
        self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.encoder2 = Unet3._block(features, features * 2, name="enc2")
        self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.encoder3 = Unet3._block(features * 2, features * 4, name="enc3")
        self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.encoder4 = Unet3._block(features * 4, features * 8, name="enc4")
        self.pool4 = nn.MaxPool2d(kernel_size=2, stride=2)

        self.bottleneck = Unet3._block(features * 8, features * 16, name="bottleneck")

        self.upconv4 = nn.ConvTranspose2d(
            features * 16, features * 8, kernel_size=2, stride=2
        )
        self.decoder4 = Unet3._block((features * 8) * 2, features * 8, name="dec4")
        self.upconv3 = nn.ConvTranspose2d(
            features * 8, features * 4, kernel_size=2, stride=2
        )
        self.decoder3 = Unet3._block((features * 4) * 2, features * 4, name="dec3")
        self.upconv2 = nn.ConvTranspose2d(
            features * 4, features * 2, kernel_size=2, stride=2
        )
        self.decoder2 = Unet3._block((features * 2) * 2, features * 2, name="dec2")
        self.upconv1 = nn.ConvTranspose2d(
            features * 2, features, kernel_size=2, stride=2
        )
        self.decoder1 = Unet3._block(features * 2, features, name="dec1")

        self.conv = nn.Conv2d(
            in_channels=features, out_channels=out_channels, kernel_size=1
        )

    def forward(self, x):
        enc1 = self.encoder1(x)
        enc2 = self.encoder2(self.pool1(enc1))
        enc3 = self.encoder3(self.pool2(enc2))
        enc4 = self.encoder4(self.pool3(enc3))

        bottleneck = self.bottleneck(self.pool4(enc4))

        dec4 = self.upconv4(bottleneck)
        dec4 = torch.cat((dec4, enc4), dim=1)
        dec4 = self.decoder4(dec4)

        dec3 = self.upconv3(dec4)
        dec3 = torch.cat((dec3, enc3), dim=1)
        dec3 = self.decoder3(dec3)

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

        dec1 = self.upconv1(dec2)
        dec1 = torch.cat((dec1, enc1), dim=1)
        dec1 = self.decoder1(dec1)

        return torch.sigmoid(self.conv(dec1))

    @staticmethod
    def _block(in_channels, features, name):
        return nn.Sequential(
            OrderedDict(
                [
                    (
                        name + "conv1",
                        nn.Conv2d(
                            in_channels=in_channels,
                            out_channels=features,
                            kernel_size=3,
                            padding=1,
                            bias=False,
                        ),
                    ),
                    (name + "norm1", nn.BatchNorm2d(num_features=features)),
                    (name + "relu1", nn.ReLU(inplace=True)),
                    (
                        name + "conv2",
                        nn.Conv2d(
                            in_channels=features,
                            out_channels=features,
                            kernel_size=3,
                            padding=1,
                            bias=False,
                        ),
                    ),
                    (name + "norm2", nn.BatchNorm2d(num_features=features)),
                    (name + "relu2", nn.ReLU(inplace=True)),
                ]
            )
        )