SegNetCMR

A Tensorflow implementation of SegNet to segments CMR images

NEW RELEASE

1. Switched to using the SELU activation function - no more batch norm and is_training hassle - a self-normalising neural network!!!!
2. To support the above - input images are rescaled from -1 to 1: (2/255.0) * image - 1.
3. Now updates the results more often and saves the checkpoint less often - this is faster. Also, doesn't flush the results after every write.

Aims

1. A demonstration of a more complete Tensorflow program including saving state and resuming.
2. Provide an ready-to-go example of medical segmentation with sufficient training and validation data, in a usable format (PNGs).

Requirements

You must have a GPU and install the tensorflow-gpu version as the cpu version does not have tf.nn.max_pool_with_argmax()

1. Python >=3.6: Best to use the Conda distribution
2. tensorflow-gpu >=0.11

Todo

1. Add code to run on your own data (currently there is only the training code present)

Running

Make sure you have conda and tensorflow installed

conda install tensorflow-gpu
python
Python 3.6.1 | packaged by conda-forge | (default, Sep  8 2016, 14:36:38)

The git clone this repository

git clone https://github.com/mshunshin/SegNetCMR.git

And start the training from the folder

cd /path/to/SegNetCMR
python train.py

And in another terminal window start tensorboard

tensorboard --logdir ./Output

Then in your webbrowser go to http://localhost:6006

Training and test data

Many thanks to the Sunnybrook Health Sciences Centre for providing a set of CMR data with associated contours. Unfortunately, in the latest release the filenames have become a little mangled, and don't match up with the contours. I have gone through the files and matched them up; exported the DICOMS as PNGs and converted the list of coordinates of the contours to PNGs as well.

The first two sets of CMRs are included as training data, the last set as test data.

With thanks to

andreaazzini/segnet: A Tensorflow SegNet translation

pydicom: A pure python dicom library

StackOverflow Tensorflow batch_norm thread

GitHub Tensorflow unpool thread

Issues and annoyances

1. The original SegNet uses max_pool_with_argmax, and requires an unpool_with_argmax. Unfortunately, Tensorflow does not provide an unpool_with_argmax. Fortunately there is code in the github thread above to make your own.
2. This version of unpool_with_argmax runs on the CPU not GPU so is a little slower.
3. Tensorflow does not provide a CPU version of max_pool_with_argmax, so if you don't have a GPU you can't run this.
4. Tensorflow forgot to include a function for gradients for maxpoolwithargmax, so it is included at the bottom of train.py
5. The name mangling of the Sunnybrook CMR data - I have fixed this and the data is included in the download.
6. SegNet works better with a version of softmax that is inversely weighted by class frequency.
7. Now using SELU as the activation funciton - this allows us to get rid of the Batch Norm (and the associated is_training hassle).

License

SegNetCMR: MIT license

SunnyBrook Cardiac Data: Public Domain

pydicom: MIT license