Xiaoyu SONG My personal Blog | Xiaoyu SONG

<< All Blogs

QuickNAT_Jan.2018

29 Oct 2018

QuickNAT : A deep fully convolution neural network (F-CNN)

  • First, apply FreeSurfer to segment scans without annotations.
  • Second, fine-tune the previous network with the limited manually annotated data.

Conclusion

  • The fully convolutional architecture offers faster processing and lager context than path-based DeepNAT.
  • The dense connections within every encoder and decoder block promote feature re-usibility in the network.

Architecture of the network

  • 4 encoders, 4 decoders separated by a bottleneck layer
  • Skip connections between all encoder and decoder blocks of the same spatial resolution
  • Each dense block consistes of 3 convolutional layers. Every convolutional layer is preceded by a batch-normalization layer and a ReLU layer. The first 2 convolutional layers are followed by a concatenation layer. (dense connections))
  • The kernel size for these 2 convolutional layers is kept small (5x5)
  • Output channels for each convolution layer is set to 64
  • The third convolutional layer has a 1x1 kernel size.
  • The classifier block is basically a convolutional layer with 1x1 kernel size, which maps the input to an N channel feature map, N is the number of class. This is followed by a soft-max layer to map the activations to probabilities.

Loss function

  • QuickNAT is learnt by optimizing 2 loss functions simultaneoulsy: weighted logistic loss; multi-class Dice loss.

Model learning

  • lr = 0.1; Reduced by one order after every 10 epochs.
  • weight decay term = 0.0001
  • Batch size = 4
  • momentum = 0.95

Training with Limited Annotated data

1. Pretraining on large unlabeled datasets with auxiliary labels

  • Use a large neuroimaging dataset: (IXI dataset), apply FreeSurfer to obtain auxiliary segmentations.
  • Pretrain QuickNAT on this large dataset with auxiliary labels.

2. Fine-tuning with limited manually labelled data

  • Use data from the Multi-Atlas Labelling Challenge dataset (Landman and Warfield,2012)
  • lr = 0.01; reduce it by an order after every 5 epochs until convergence