The ImFusion Suite provides an Cone-Beam Computed Tomography (CBCT) by offering state of the art reconstruction techniques.
Independently of the actual reconstruction method the X-Ray projections can be pre-processed, e.g. log-converted and/or filtered, prior to passing them to the reconstruction step. With respect to the actual reconstruction we have the two main methods for CBCT reconstruction available, namely filtered backprojection via the FDK algorithm, as well as iterative reconstruction with an ordered subset (OS) approach. Following the reconstruction step one may apply post-processing methods to the resulting volumetric representation of the X-Ray attenuation. Below we will give a more in-depth description of the supported methods for each of these three steps.
The X-Ray projection images are first loaded and pre-processed, which includes extrapolation of dead mask pixels, gain correction by applying an air image and one or multiple gain maps, and subsequent log-conversion. For iterative reconstruction, many 2D image filters in a fast GPU implementation can be applied before or after the log-conversion (sharpening, smoothing, median filter, anisotropic, diffusion, bilateral filter, among others).
Tomographic reconstruction of X-Ray projections in general evolves around the simulation of X-Ray projections, also known as projection, and the back-projection of an X-Ray projection value along the corresponding X-Ray from source to detector pixel. While the FDK purely relies on the latter, iterative approaches use a combination of these two components in order to iteratively update the estimated tomographic reconstruction.
This algorithm works according to the 3D cone-beam reconstruction algorithm proposed by Feldkamp, Davis, Kress (FDK). A Ramp-Filter is first applied to the projection images with a number of possible settings (Ram-Lak, Hamming, Cosine, Shepp-Logan). Further cosine weightings are applied as required, and if a short-scan is present, Parker weights are considered. Both, the filtering as well as the subsequent back-projection algorithm are executed efficiently on the GPU which enables to rapidly create the reconstruction volume.
For iterative reconstruction, we feature the commonly known simultaneous iterative reconstruction technique (SIRT) as well as a maximum likelihood expectation-maximization (MLEM) approach. In both cases we apply an ordered subsets (OS) scheme, i.e. instead of using the complete set of all X-Ray projections, in each iteration only a subset is used. Within the ImFusion Suite these subsets consist of 1 to 30 projections at a time. Each of these subsets are forward and back-projected using fast GPU-based operations. Additionally we support region-of-interest (ROI) reconstruction, such that interior reconstruction problems can be solved. Within this approach scans where parts of the subject is outside of the projection can be reconstruction more reliably despite the missing information. Optionally, a total variation (TV) based regularization can be enabled which proved to be a valuable regularization for CT reconstruction for e.g. low-dose scenarios.
After one of the two reconstruction methods described above completed, the volume may be further processed using a number of 3D image filters (sharpening, smoothing, anisotropic diffusion). Then values outside the projection overlap are erased to zero, the intensities are clamped to a valid range for export, and the reconstruction volume is downloaded to CPU memory with the desired bit precision. A DICOM export then stores the volume, while defining the tags rescale slope/intercept tags properly to arrive at Hounsfield values.