ctImage

This is a repository for the Learning with context encoding for single-stage cranial bone labeling and landmark localization. This repository contains the final model as described in the manuscript. The model is stored as a TorchScript jit file MiccalFinalModel.dat.

Dependencies:

• Python

• Pytorch

• NumPy

• SimpleITK

• VTK

• scipy

• skimage

• TorchIO

  Once Python is installed, each of these packages can be downloaded using Python pip

Using the code

Due to data privacy agreements, we are not able to share any example CT Image. Our example inference codes are based on the data processing script to generate masked normalzied CT images based on MHA graphic data files (.mha). If you are not using the same data format for the CT images, you could generate input array based on your own choice. Our model requires input image arrays with size of 96x96x96 and intensity normalized to the range of 0-1.

Quick summary

Input: MHA graphic files.

Output: MHA graphic files labeling 5 cranial bone, and VTK PolyData containing 4 landmarks at the cranial base.

Code example

An example for automatic concurrent cranial bone labeling and landmark localization is:

import DataProcessing
import SimpleITK as sitk
import ModelConfiguration

### process example CT image

ctImage = sitk.ReadImage('./ExampleCTImage.mha')
binaryImage = DataProcessing.CreateBoneMask(ctImage)
ctImage = DataProcessing.ResampleAndMaskImage(ctImage, binaryImage)

### model
modelPath = './MiccaiFinalModel.dat'
device = ModelConfiguration.getDevice()
model = ModelConfiguration.adaptModel(modelPath, device)
imageData = ModelConfiguration.adaptData(ctImage, device)

landmarks, boneLabels = ModelConfiguration.runModel(model, ctImage, binaryImage, imageData)

When using this code, be sure to assign corret path containing a valid MHA graphic file of a CT image to the ctImage.

The workflow

• The CreateBoneMask function creates a binary mask for the CT image.
• The ResampleAndMaskImage function resamples and masks the CT image to the correct size and normalizes the intensity.
• The getDevice function specifies the device for the torch model.
• The getDevice function specifies the device for the torch model.
• The adaptModel and adaptData functions prepares the model and data for inference.
• The runModel function generates landmark and bone labeling predictions and resample them to the original CT image space.

If you have any questions, please email Jiawei Liu at jiawei.liu@cuanschutz.edu