Learning Day 70: 3D U-Net with 3D convolution layers, V-Net, DenseNet, FC-DenseNet
Published in
3 min readJun 25, 2021
Background
- Continue from Day 69 regarding U-Net, one problem with biomedical imaging is that different slices at different locations would appear very differently, thus U-Net is not able to identify the target region
- What if we directly feed the 3D data into a U-Net?
3D U-Net
- It employs 3D convolution layers to take in 3-dimension imaging data
- Besides 3D conv layers, the principle of 3D U-Net is very similar to U-Net
3D convolution
- It is very similar to 2D convolution, just that we need to imagine 1 more dimension
- A 2D convolution operation
- A 3D convolution operation with a 3D input of 1 input channel
- A 3D convolution operation with 3D input of input channel, c = 3. In this case, the input is actually 4-dimension. The filter is 5-dimension.
V-Net
- Added ResNet-like shortcut
- Same loss function (DICE coefficient)
- Include “down” conv layer (conv with stride=2)
- Make use of 1x1x1 conv layer prior to softmax for classification between background and foreground
- PReLU activation where the gradient at y<0, a, is learnable
DenseNet
- Focus on fully utilising the feature maps
- Reduce the no. of parameters
- Not prone to overfitting even for small dataset
- There are extra connections between all layers. I.e. Current layer’s input includes outputs from all previous layer
- Eg. Input X0 →Output H1, Input X0 &X1 →Output H2
FC-DenseNet
- The architecture is still similar to U-Net
- Just replace the original conv blocks withe dense block
- Skip connections exist only at the scaling down side (left), not at the scaling up side (right) of the network
- Skip connections are to avoid feature loss
Reference
Unless otherwise stated, all presented contents are summarised from this course for personal learning purposes