Synthetic X-ray CONSTRUCTION
Motivation
To create a neural-network that could receive X-rays as input and produce CT-quality 3D radiodensities as output, the training dataset have to be consisted of paired 2D-3D data.
A matching X-Ray - CT volume is difficult to achieve since they are two distinctive modalities, and that patient position would not be exactly the same. Therefore we have decided to create synthetic X-ray images from CT scans using ray-tracing algorithm, serving as the matching 2D-3D data in our training dataset.
We have selected the Siddon-Jacobs ray-tracing algorithm to calculate the radiological path traversed through a reconstructed CT volume from a point-source projection approach. Among the literature we have referred to, only Moturu & Chang (2019) and Henzler et al. (2018) have generated synthetic X-ray images, but they did it using a parallel-ray perspective, which is far from actual X-ray creation process.
CREATING SYNTHETIC X-RAY IMAGES
3D CT Volume
3D Volumetric dataset in CT scans are reconstructed by stacking cross sections (axial images) in the Z-direction. Air and other unnecessary voxels are removed, leaving only the patient body in the reconstructed volume.
Each reconstructed CT volume are then used as input to a pipeline setup to mirror the process of X-ray imaging (see following).
Ray-Tracing
Using Siddon-Jacobs ray-tracing algorithm, we can simulate the process of X-ray by propagating the incident X-ray photons (from radiation source) through the reconstructed volume's 3D voxels (see bottom image on left). Top image in the left panel shows an actual X-ray procedure.
Some example parameters involved in this process are: rotation / translation of body (input volume), distance from radiation source to body and distance from body to image.
The resulting image is then post-processed with adaptive histogram equalization for enhancement.
The Siddon-Jacobs algorithm is available as a module in the well-known Insight Segmentation and Registration toolkit (ITK) imaging package.
Example
A synthetic X-ray image after ray-tracing and adaptive histogram equalization (left).
Compared to the SCOUT image of the same patient (right), our synthetic X-ray image is proven to be very similar to actual X-ray. Visual differences in brightness/contrast are largely due to different windowing/threshold levels for each image. While the synthetic x-ray might look darker than the SCOUT image, this is not reflected in the underlying data.
Note: SCOUT scans are pilot images taken for positional purposes prior to a CT imaging study. They are essentially X-rays but are excluded from consideration to be used as 2D input or paired data in our project as they do not provide a complete match to our CT reconstructed volumes.
Synthetic X-Ray
