3dseg for CT-Data

This reposirory is based on pytorch-3dunet implementation.

TODO's

• create bigger dataset of volume

💿 Installation

🔥 PyTorch

This reposirotry is builded with

macOS

Please not also that 3D-segemtantion is not available on a macOS machine, as pytorch with metal support only supports 4D arrays!

Clone this repository using the terminal:

git clone https://github.com/choROPeNt/3dseg.git

and navigate to it in your terminal.

cd 3dseg

Then run:

python -m pip install -e .

This should install the 3dseg python package via PIP in the current active virtual enviroment. How to set up a virtual enviroment please refer to virtual enviroment section

3DUnet model

The model is a 3D-UNet

losses

Dice Loss

The DiceLoss $\mathcal{L}{Dice}$ is defined as following $$\mathcal{L}{\text{Dice}} = \frac{2 \sum_i^N p_i g_i}{\sum_i^Np_i^2+\sum_i^N g_i^2}$$

binary-corss entropy

• Pytorch - BCEloss

$$ \mathcal{L}_{\text{BCE}}(x,y) = { l_1,\dots,l_N}^\top \quad l_N=-w_n \left[ y_n ; \log x_n + (1-y_N);\log (1-x_n)\right]$$

$\mathcal{L}_{\text{BCE}}$ over one batch is determined by

$$\mathcal{L}_{\text{BCE}}(x,y) = \left{ \begin{array}{ll} \frac{1}{N} \sum_i^N l_i & \text{if reduction = 'mean'} \\ \sum_i^N l_i & \text{if reduction = 'sum'} \\ \end{array} \right.$$

BCE-DiceLoss

linear combination of BCE and Dice loss

$$ $$

training

python ./scripts/train.py --config <CONFIG>

[b,c,d,w,h] -> [b,3*c,d,w,h]

support

virtual enviroment

to create a virtual enviroment named .venv you may run

python -m venv .venv

to activate the virtual enviroment please run

• for linux and macos

source .venv/bin/activate

Support and Advanced Installations

HDF5 with MPI (macOS)

If you are using a macOS machine and want to access your results faster, it's reccommended to install HDF5 with MPI support.

make sure you have installed HDF5 with MPI support. Otherwise you can install it using the Homebrew formula

brew install hdf5-mpi

After a succefull installation you have to set the PATH variables for the compiler

export CC=mpicc
export HDF5_MPI="ON"
export HDF5_DIR="/path/to/parallel/hdf5"  # If this isn't found by default

If you are unsure where the location of your HDF5-MPI binary is, you can find the HDF5_DIR with running

h5pcc -showconfig

In the case you have an

Install the python package h5py with

 % python -m pip install --no-binary=h5py --no-cache-dir h5py

the options --no-binary=h5py and --no-cache-dir that pip is forced to build the package from source via wheel