/Feature-interaction-network-based-on-Hierarchical-Decoupled-Convolution

Feature interaction network based on Hierarchical Decoupled Convolution for 3D medical image segmentation

Primary LanguagePython

Feature interaction network based on Hierarchical Decoupled Convolution for 3D medical image segmentation

This repository is the work of "Feature interaction network based on Hierarchical Decoupled Convolution for 3D medical image segmentation" based on PyTorch implementation.

Requirements

  • python 3.6
  • pytorch 0.4 or 1.0
  • nibabel
  • pickle
  • imageio
  • pyyaml

Implementation

Download the BraTS2019 dataset and change the path:
experiments/PATH.yaml

Data preprocess:(Convert the .nii files as .pkl files. Normalization with zero-mean and unit variance)
python preprocess.py

(Optional) Split the training set into k-fold for the cross-validation experiment.
python split.py

Training

Sync bacth normalization is used so that a proper batch size is important to obtain a decent performance. Multiply gpus training with batch_size=8 is recommended.
python train_all.py --gpu=0,1 --cfg=FHDC_Net --batch_size=8

Test

You could obtain the resutls as paper reported by running the following code:
python test.py --mode=1 --is_out=True --verbose=True --use_TTA=True --postprocess=True --snapshot=True --restore=model_last.pth --cfg=FHDC_Net --gpu=1

Evaluation

Submit the results to the online evaluation server.