3D_DenseSeg: 3D Densely Convolutional Networks for Volumetric Segmentation

By Toan Duc Bui, Jitae Shin, Taesup Moon

This is the implementation of our method in the MICCAI Grand Challenge on 6-month infant brain MRI segmentation-in conjunction with MICCAI 2017.

Update: (Aug. 14. 2019): We also release the Pytorch version of my journal version at https://github.com/tbuikr/3D-SkipDenseSeg

Citation

@article{bui2019skip,
  title={Skip-connected 3D DenseNet for volumetric infant brain MRI segmentation},
  author={Bui, Toan Duc and Shin, Jitae and Moon, Taesup},
  journal={Biomedical Signal Processing and Control},
  volume={54},
  pages={101613},
  year={2019},
  publisher={Elsevier}
}

Link journal version: https://www.sciencedirect.com/science/article/pii/S1746809419301946

Link conference version: https://arxiv.org/abs/1709.03199

Introduction

6-month infant brain MRI segmentation aims to segment the brain into: White matter, Gray matter, and Cerebrospinal fluid. It is a difficult task due to larger overlapping between tissues, low contrast intensity. We treat the problem by using very deep 3D convolution neural network. Our result achieved the top performance in 6 performance metrics.

Dice Coefficient (DC) for 9th subject

	CSF	GM	WM	Average
3D-DenseSeg	94.74%	91.61%	91.30%	92.55%

Citation

@article{bui20173d,
  title={3D Densely Convolution Networks for Volumetric Segmentation},
  author={Bui, Toan Duc and Shin, Jitae and Moon, Taesup},
  journal={arXiv preprint arXiv:1709.03199},
  year={2017}
}

Requirements:

• 3D-CAFFE (as below), python 2.7, Ubuntu 14.04, CUDNN 5.1, CUDA 8.0
• TiTan X Pascal 12GB

Installation

• Step 1: Download the source code

git clone https://github.com/tbuikr/3D_DenseSeg.git
cd 3D_DenseSeg

• Step 2: Download dataset at http://iseg2017.web.unc.edu/download/ and change the path of the dataset data_path and saved path target_path in file prepare_hdf5_cutedge.py

data_path = '/path/to/your/dataset/'
target_path = '/path/to/your/save/hdf5 folder/'

• Step 3: Generate hdf5 dataset

python prepare_hdf5_cutedge.py

• Step 4: Run training

./run_train.sh

• Step 5: Generate score map and segmentation image. You have to change the path in the file seg_deploy.py as

caffe_root = '/path/to/your/caffe/build/tools/caffe/'# (i.e '/home/toanhoi/caffe/build/tools/caffe/')

And run

python seg_deploy.py

3D CAFFE

For CAFFE, we use 3D UNet CAFFE with minor modification. Hence, you first download the 3D UNet CAFFE at

https://lmb.informatik.uni-freiburg.de/resources/opensource/unet.en.html

And run the installation as the README file. Then we change the HDF5DataLayer that allows to randomly crop patch based on the code at https://github.com/yulequan/3D-Caffe You can download the code by

git clone https://github.com/yulequan/3D-Caffe/
cd 3D-Caffe
git checkout 3D-Caffe
cd ../

After downloading both source codes, we have two folder code 3D-Caffe and caffe (for 3D UNet CAFFE). We have to copy the hdf5 data files from 3D-Caffe to caffe by the commands

cp ./3D-Caffe/src/caffe/layers/hdf5_data_layer.cpp ./caffe/src/caffe/layers/
cp ./3D-Caffe/src/caffe/layers/hdf5_data_layer.cu ./caffe/src/caffe/layers/
cp ./3D-Caffe/include/caffe/layers/hdf5_data_layer.hpp ./caffe/include/caffe/layers/hdf5_data_layer.hpp

Then add these lines in the field message TransformationParameter of the file caffe.proto in the ./caffe/src/caffe/proto (3D UNet CAFFE)

optional uint32 crop_size_w = 8 [default = 0];
optional uint32 crop_size_h = 9 [default = 0];
optional uint32 crop_size_l = 10 [default = 0];

Add following code in the ./caffe/include/caffe/filler.hpp

/**
  3D bilinear filler 
*/
template <typename Dtype>
class BilinearFiller_3D : public Filler<Dtype> {
 public:
  explicit BilinearFiller_3D(const FillerParameter& param)
      : Filler<Dtype>(param) {}
  virtual void Fill(Blob<Dtype>* blob) {
    CHECK_EQ(blob->num_axes(), 5) << "Blob must be 5 dim.";
    CHECK_EQ(blob->shape(-1), blob->shape(-2)) << "Filter must be square";
    CHECK_EQ(blob->shape(-2), blob->shape(-3)) << "Filter must be square";
    Dtype* data = blob->mutable_cpu_data();

    int f = ceil(blob->shape(-1) / 2.);
    float c = (2 * f - 1 - f % 2) / (2. * f);
    for (int i = 0; i < blob->count(); ++i) {
      float x = i % blob->shape(-1);
      float y = (i / blob->shape(-1)) % blob->shape(-2);
      float z = (i/(blob->shape(-1)*blob->shape(-2))) % blob->shape(-3);
      data[i] = (1 - fabs(x / f - c)) * (1 - fabs(y / f - c)) * (1-fabs(z / f - c));
    }
    

    CHECK_EQ(this->filler_param_.sparse(), -1)
         << "Sparsity not supported by this Filler.";
  }
};

and in the GetFiller(const FillerParameter& param) function (same file)

else if (type == "bilinear_3D"){
    return new BilinearFiller_3D<Dtype>(param);
  }

Final, you recompile 3D UNet CAFFE (uncomment USE_CUDNN := 1) and can you my prototxt. Please cite these papers when you use the CAFFE code

Note

• If you want to generate network prototxt, you have to change the path of caffe_root

caffe_root = '/path/to/your/caffe/build/tools/caffe/'# (i.e '/home/toanhoi/caffe/build/tools/caffe/')

And run

python make_3D_DenseSeg.py

• If you have the error AttributeError: 'LayerParameter' object has no attribute 'shuffle' when run python make_3D_DenseSeg.py, then you can fix it by replacing the line 35 in the net_spec.py:

  #param_names = [s for s in dir(layer) if s.endswith('_param')]
  param_names = [f.name for f in layer.DESCRIPTOR.fields if f.name.endswith('_param')]

• Plot training loss during training

python plot_trainingloss.py ./log/3D_DenseSeg.log