AIGT

This repository contains scripts for deep learning for guided medical interventions. For some projects, the complete workflow is implemented, from formatting and annotations to deployment of models in real time. Most projects use ultrasound imaging.

Getting started

Install and set up Anaconda environment

Install Anaconda (with Python 3.7)
Clone this repository on your computer. Your local clone may be in a path like c:\dev\aigt
Start the Anaconda Prompt application and navigate to the environment setup folder cd c:\dev\aigt\SetupAnaconda
Run the setup_env.bat file to create environment in a folder, e.g. setup_env.bat c:\dev\dlenv This will install TensorFlow 2.0 and other packages that are used by projects. The previous environment setup script (for TensorFlow v1 is still available as setup_env_tf1.bat

Additional local files you may need, but they are not in the code repository

Please do not commit/push these local files, as everybody sets them up with values that only apply to their environment.

local_vars.py - Some notebooks require a file in the Notebooks folder of your local repository clone, named local_vars.py. This file should define the root_folder variable. The file may just contain this single line of code: root_folder = r"c:\Data".
girder_apikey_read.py - Some notebooks require a file named girder_apikey_read.py with a single line of code that specifies your personal API key to the private data collection. If you work with non-public data stored on a Girder server, ask your supervisor for a Girder account and how to generate API keys for yourself.

To run Slicer notebooks

Install Slicer 4.11 or newer version (later than 2019-09-16 is recommended, for full functionality)
Install the SlicerJupyter extension for Slicer, and follow the extension user guide to add Slicer as a kernel in Jupyter Notebook (use the Copy command to clipboard button and paste it in the active Anaconda environment).
If you have a GPU and would like Slicer's TensorFlow to use it, then install CUDA 10.1 and cuDNN 7.6.5. GPUs can make training of models much faster, but may not significantly speed up trained models for prediction compared to CPUs.
Some users have reported that they needed to install Visual Studio (2015 or later) to be able to use TensorFlow.
Install additional packages in the Slicer python environment, to be able to run all Slicer notebooks. Use the Python console of Slicer to run this command:

pip_install("tensorflow opencv-contrib-python girder_client pandas nbformat nbconvert")

To run notebooks, start the Anaconda command prompt, navigate to the aigt folder, and type the jupyter notebook command.

Acquire tracked ultrasound

Use the Sequences extension in 3D Slicer to record tracked ultrasound sequences.
Record Image_Image (the ultrasound in the Image coordinate system) and ImageToReference transform sequences. Note that Slicer cannot record transformed images, so recording Image_Reference is not an option.
If you work on segmentation, you can use the SingleSliceSegmentation Slicer module in this repository to speed up manual segmentation.
Create annotations by placing fiducials or creating segmentations with the SingleSliceSegmentation module.
For segmentations, you may use the SingleSliceSegmentation module to export images and segmentations in png file format.
Use scripts in the Notebooks/Slicer folder to export images and annotations.

Process exported data

It is recommended that you save separate sequences for validation and testing
Use Notebooks/FoldersToSavedArrays to save data sequences as single files for faster data loading during training.