|
|
import torch.nn as nn |
|
|
from monai.networks.nets import resnet101, resnet50, resnet18, ViT |
|
|
import torch |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
class ResNet50_3D(nn.Module): |
|
|
def __init__(self): |
|
|
super(ResNet50_3D, self).__init__() |
|
|
|
|
|
resnet = resnet50(pretrained=False) |
|
|
resnet.conv1 = nn.Conv3d(1, 64, kernel_size=7, stride=2, padding=3, bias=False) |
|
|
hidden_dim = resnet.fc.in_features |
|
|
self.backbone = resnet |
|
|
self.backbone.fc = nn.Identity() |
|
|
|
|
|
def forward(self, x): |
|
|
x = self.backbone(x) |
|
|
return x |
|
|
|
|
|
|
|
|
|
|
|
class Classifier(nn.Module): |
|
|
""" Classifier class with FC layer and single output neuron """ |
|
|
def __init__(self, d_model, hidden_dim=1024, num_classes=1): |
|
|
super(Classifier, self).__init__() |
|
|
self.fc = nn.Linear(d_model, num_classes) |
|
|
def forward(self, x): |
|
|
x = self.fc(x) |
|
|
return x |
|
|
|
|
|
|
|
|
class Backbone(nn.Module): |
|
|
""" ResNet 3D Backbone""" |
|
|
|
|
|
def __init__(self): |
|
|
super(Backbone, self).__init__() |
|
|
|
|
|
resnet = resnet50(pretrained=False) |
|
|
resnet.conv1 = nn.Conv3d(1, 64, kernel_size=7, stride=2, padding=3, bias=False) |
|
|
hidden_dim = resnet.fc.in_features |
|
|
self.backbone = resnet |
|
|
self.backbone.fc = nn.Identity() |
|
|
|
|
|
def forward(self, x): |
|
|
x = self.backbone(x) |
|
|
return x |
|
|
|
|
|
|
|
|
class SingleScanModel(nn.Module): |
|
|
""" End to end model with backbone and classifier""" |
|
|
|
|
|
def __init__(self, backbone, classifier): |
|
|
super(SingleScanModel, self).__init__() |
|
|
self.backbone = backbone |
|
|
self.classifier = classifier |
|
|
self.dropout = nn.Dropout(p=0.2) |
|
|
|
|
|
|
|
|
def forward(self, x): |
|
|
|
|
|
x = self.backbone(x) |
|
|
x = self.dropout(x) |
|
|
x = self.classifier(x) |
|
|
return x |
|
|
|
|
|
class SingleScanModelBP(nn.Module): |
|
|
""" End to end model with backbone and classifier that takes 2 input scans at once""" |
|
|
|
|
|
def __init__(self, backbone, classifier): |
|
|
super(SingleScanModelBP, self).__init__() |
|
|
self.backbone = backbone |
|
|
self.classifier = classifier |
|
|
self.dropout = nn.Dropout(p=0.2) |
|
|
self.bilinear_pooling = nn.Bilinear(in1_features=2048, in2_features=2048, out_features=512) |
|
|
|
|
|
|
|
|
def forward(self, x): |
|
|
x = [self.backbone(scan) for scan in x.split(1, dim=1)] |
|
|
features = torch.stack(x, dim=1).squeeze(2) |
|
|
merged_features = torch.mean(features, dim=1) |
|
|
merged_features = self.dropout(merged_features) |
|
|
output = self.classifier(merged_features) |
|
|
return output |