Fig 1.
An overview of the SPIR technique.
In the first of two steps, the projection data is decomposed into several basis functions from which the pseudo-monochromatic projection is calculated and the metal image is reconstructed. Using these as input and prior respectively, the image is iteratively reconstructed in the second step, while at the same time knowledge on the location and density of the metal implant is enforced and corrected.
Fig 2.
One of the virtual phantoms used in this work.
It contains several anatomic components of a jaw such as soft tissue, spine, bone marrow and teeth. A metal implant is embedded on one of the teeth. In this work, we simulated three different shapes of metal implants: circle, horseshoe and triangle.
Fig 3.
The X-ray spectrum used in the simulation.
The x-ray source was generated at tube voltage of 125 kV with Wolfram target and aluminum filter of thickness 2.7mm, yielding mean spectrum energy of 55.457 keV.
Fig 4.
The original model of the phantom and the images of the decomposed basis functions reconstructed with filtered back-projection (FBP).
The second column shows photoelectric attenuation, third column the Compton scattering, and the fourth column the gold attenuation. Row-wise are the different shapes of the metal implant: first row circle, second row horseshoe, and bottom row triangle. It can be seen that, the location of the gold implant is accurately detected, while the gold implant can be distinguished from other parts of the phantom, especially the teeth. The reconstructed images are normalized to 1, and have WW of 0.2 and WL of 0.1.
Fig 5.
The zoom-in of the metal implants.
The zoom-in of the metal implants obtained shown in the top-row indicates the accuracy of the material decomposition technique in detecting the metal implant. When compared to the original model, only a slight difference as in 1 pixel is observed, which can be attributed to discretization error. The average difference in density is about 0.285 g/cm3. The top-row images are normalized to 1, with WW of 1.0 and WL of 0.5. The bottom-row images have WW 1.0 and WL of -0.5.
Fig 6.
The reconstructions of the phantom using different algorithms.
The reconstructions of the first row is done using FBP, second row penalized maximum likelihood iterative reconstruction without prior (IR), and the third row Spectral-driven Iterative Reconstruction (SPIR). Column wise are the different shapes of metal implant at full-view and zoomed-view. The first column-pair has the shape circle, second column-pair horseshoe, and third column-pair triangle. All images have WW of 1000 HU and WL of 4000 HU.
Fig 7.
The vertical line profiles as marked in Fig 4 for different reconstruction algorithms.
The line profile from SPIR algorithm reflects the true attenuation values best, in comparison to FBP and IR.
Fig 8.
The influence of the prior information on the outcome of the model-based iterative reconstruction.
The results with a smaller (A and B) or larger (D and E) metal implant show that inaccurate prior results in a less than optimal image. The reconstruction result with the original model is shown in C for comparison. All images have WW of 4000 HU and WL of 1000 HU.