highresnet

$ NII_FILE=`download_oasis`
$ deepgif $NII_FILE

PyTorch implementation of HighRes3DNet from Li et al. 2017, *On the Compactness, Efficiency, and Representation of 3D Convolutional Networks: Brain Parcellation as a Pretext Task*.

All the information about how the weights were ported from NiftyNet can be found in my submission to the MICCAI Educational Challenge 2019.

Usage

Command line interface

(deepgif) $ deepgif t1_mri.nii.gz
Using cache found in /home/fernando/.cache/torch/hub/fepegar_highresnet_master
100%|███████████████████████████████████████████| 36/36 [01:13<00:00,  2.05s/it]

PyTorch Hub

If you are using pytorch>=1.1.0, you can import the model directly from this repository using PyTorch Hub.

import torch
repo = 'fepegar/highresnet'
model_name = 'highres3dnet'
print(torch.hub.help(repo, model_name))
"HighRes3DNet by Li et al. 2017 for T1-MRI brain parcellation"
"pretrained (bool): load parameters from pretrained model"
model = torch.hub.load(repo, model_name, pretrained=True)

Installation

1. Create a conda environment (recommended)

ENVNAME="gifenv"  # for example
conda create -n $ENVNAME python -y
conda activate $ENVNAME

2. Install PyTorch and highresnet

Within the conda environment:

pip install light-the-torch  # to get the best PyTorch
ltt install torch  # to get the best PyTorch
pip install highresnet

Now you can do

from highresnet import HighRes3DNet
model = HighRes3DNet(in_channels=1, out_channels=160)