Code repository of the Young eScientist Award Project with the 3D Innovation Lab at the VUmc. This code accompanies a paper titled:
"CT image segmentation of bone for medical additive manufacturing using convolutional neural networks"
which is currently under review.
Bone segmentation of CT scans is an essential step during medical treatment planning. The exact thickness, orientation and position of bony structures is required to make patient-specific constructs such as surgical guides and implants. During bone segmentation each pixel in a medical image is classified as either 'bone' or 'background'. Unfortunately, current algorithms either lack robustness and reliability, or require tedious manual interaction (Van Eijnatten et al., 2018). Therefore, this repository contains a fully-automated convolutional neural network (CNN) to perform bone segmentation of CT scans.
The CNN was trained using CT scans of three patients that were previously treated at the Vrije Universiteit Medical Center. Each pixel of the CT scans was labelled as either "bone" or "background" based on the knowledge of highly experienced medical engineers. A number of 500,000 pixels was randomly selected to create a axial patch of 33x33 around those selected pixels. These patches were subsequently used to train the CNN, which allow the CNN to learn the classification of pixels based on the local environment. Furthermore, a balanced training method was applied as proposed by Havaei et al. (2016), which implies that "bone" and "background" patches were equally represented in the dataset.
Our CNN architecture was inspired by the one described in the Github repository of N.Aldenborgh. This architecture consists of four convolutional layers with additional ReLU, batch normalization and pooling layers. Pixel classification was performed using the softmax cost function. Furthermore, the RMSprop optimizer was used during the CNN training.
Clone the repository:
sudo apt-get install git
git clone https://github.com/NLeSC/yeap16-ai-3d-printing.gitCreate an environment with all dependencies:
conda env create -f environment.yml
source activate deepy3dInstall the package:
python setup.py installWhen called directly, the Deepy3d class runs an example script that reads in patient data according to a configuration file, trains a CNN model and visualizes its predictions on a held-out patient. For patient confidentiality reasons, we cannot include our CT data set in this repository. But we included a config.yml and ConfigReader class to facilitate reading in your own patient data.
In order to run the example script:
- Insert your local data directories in config.yml.
- Call Deepy3d class:
python ./deepy3d/Deepy3d.pyconfig.yml also contains network architecture and training parameters, should you want to change those as well.
The repository also contains a fork of a medical image processing and visualization package called 3D-Slicer. We used Slicer for manipulating CT-scans and performing global thresholding on the grayscale intensity values (Hounsfield units).
Install dependencies (Ubuntu):
sudo apt-get -y install unzip curl wget xz-utils bsdtarCall Slicer installation script:
sh -c "./slicer/installSlicer.sh)"Download sample data from slicer.org:
./tools/getSampleData.bashApply thresholding:
./tools/launchSlicer.sh examples/slicer_CT_threshold.pyBugs, questions and comments can be submitted in the issues tracker.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent