Nifty Assignment: Training a neural network for medical image processing

Background

Medical imaging is becoming an increasingly popular application of Machine Learning. Medical practitioners can use software to obtain diagnosis or second opinions on X-Ray images, lowering the chances of a missed threat.

In this assignment, students will be able to set up a model to train using the Deep Learning Tool Kit and Tensorflow.

Meta-information

Attribute	Explanation
Summary	Set-up an environment with TensorFlow and the Deep Learning Tool Kit to train and run inference on sample medical imagesets.
Topics	Machine Learning, Neural Networks, TensorFlow.
Audience	CS0, CS1 to get their feet wet and introduce concepts, CS2+ for development.
Difficulty	Easy to set up and get running, with little room for error.
Strengths	The strength of this assignment is being able to create a training environment quickly.
Weaknesses	The topics that the Deep Learning Tool Kit and TensorFlow require are hard to digest and understand.
Dependencies	Requires understanding of Machine Learning training models and understanding of visual data.
Variants	Advanced Machine Learning classes may delve into the source code of the training models.

System Requirements

GNU/Linux system
- Python 2 installed - DLTK highly recommends Python 3, but you may come across erros when installing tensorflow-gpu
  - Pip is installed
  - virtualenv is installed
NVIDIA GPU for faster training that supports CUDA
Plenty of hard drive space, as DLTK uses training sets large in file size

Assignment instructions

Setting up DLTK

Clone the DLTK git repository

git clone https://github.com/DLTK/DLTK.git

Navigate to the DLTK folder
```
cd DLTK
```
Create a virtual python environment
```
virtualenv -p python2 venv_tf
```
Activate the virtual environment
```
source venv_tf/bin/activate
```
Install the dependencies for DLTK within the virtual environment
```
pip install -e .
```
Install tensorflow
```
pip install tensorflow-gpu==1.4.0
```

Downloading the IXI data sets for training

For more information on the IXI dataset, visit https://brain-development.org/ixi-dataset/

Go to the IXI_HH folder under data
```
cd DLTK/data/IXI_HH
```
Run the download script
```
python download_IXI_HH.py
```
Note that the download may take a long time and may need to be left overnight depending on your internet connetion.

Start training

IXI_HH sex classification using resnet

Note: Any time you are training or inferencing, make sure the python virtual environment is enabled

   source DLTK/venv_tf/bin/activate

Go to the IXI_HH_sex_classification folder

cd DLTK/examples/applications/IXI_HH_sex_classification

Create a folder to keep your model progress in
```
mkdir my_model
```

Begin training

python train.py --verbose --cuda_devices 0 --model_path my_model

Start up TensorBoard to observe the training progress
```
tensorboard --logdir .
```
Open a browser to access Tensorboard, by default, Tensorboard opens on port 6006
```
http://localhost:6006
```

Tensorboard interface

Accuracy and Loss graphs

Accurcy and Loss

Image evaluation

eval

t-SNE graphs

For information on how to effectively understand t-SNE graphs, see the following Distill entry: https://distill.pub/2016/misread-tsne/

tsne

Inference

By default, inference runs on random images from the data set and infers whether they are male or female.

   python -u deploy.py --model_path my_model

inference

Credits

Assignment possible thanks to the work of the great people at the Deep Learning Tool Kit