scouvreur / wholeheartmrisegmenter

Whole Heart MRI Segmenter based on data from HVSMR MICCAI 2016 Challenge

Whole Heart MRI Segmenter

This data is from the HVSMR 2016: MICCAI Workshop on Whole-Heart and Great Vessel Segmentation from 3D Cardiovascular MRI in Congenital Heart Disease challenge.

This segmenting command-line tools is based on transfer learning from a Dense VNet application in NiftyNet trained on CT data for abdominal organ segmentation.

If you want to make any changes or improvements to the config file, you can use the NiftyNet config file docs, or the complete docs. You can also find the source code for NiftyNet on GitHub along with other sample config files to use.

Setup

First, clone this repo using:

>>> git clone https://github.com/scouvreur/WholeHeartMRISegmenter.git

Then, do:

>>> cd WholeHeartMRISegmenter/
>>> conda create -n niftynet-dev
>>> conda activate niftynet-dev
>>> conda install tensorflow-gpu==1.12
>>> pip install niftynet opencv-python scikit-image simpleitk pyyaml

Data and Labels

Labels preview

Label	Axial	Coronal	Sagittal	Classes
n/a				num_classes=1
A				num_classes=3
B				num_classes=5

Labels explanation in ITKSnap

Label	3D View	ITKSnap Labels	Anatomical Explanation
A			Label 1: Myocardium Label 2: Blood Pool and Great Vessels
B			Label 1: Myocardium Label 2: Blood Pool Label 64: Aorta Label 71: Pulmonary Arteries

Training

To start training, within the folder use:

>>> net_segment train -c config.ini

It is recommended to use around max_iter = 2000 to reach good results.

Inference

To perform inference on the data in testing/, you can use:

>>> net_segment inference -c config.ini

You can select the iteration checkpoint from training to perform inference on using inference_iter. Outputted segmentations will be placed in the directory specified in save_seg_dir, with the original name and the posfix specified in output_postfix.

wholeheartmrisegmenter's Issues

Short-axis, cropped training data

Hello! I am a postgraduate and major in medical image analysis. I wanted to evaluate some methods on HVSMR but found that this dataset had been removed from the website. I wonder if you have the short-axis, cropped training data? Thank you very much!