Introduction C-arm based Cone-Beam CT (CBCT) imaging enables the in-situ acquisition of three dimensional images. In the context of image-guided interventions this technology has the potential to reduce the complexity of a procedure's workflow, with imaging and intervention carried out in the same location. For interventions carried out in the thoracic-abdominal regions, image acquisition while the patient is freely respiring results in blurred images. To improve image quality, data is acquired either at breath-hold or using retrospective gating approaches, resulting in a 4D (space+time) reconstruction. To acquire a 4D CBCT image using retrospective gating, the acquired projection X-ray images are grouped into separate data sets based on their phase in the respiratory cycle. Standard reconstruction is then performed for each data set. Motion artifacts are reduced as the number of data sets is increased. On the other hand this also reduces the number of images in each data set which increases the reconstruction artifacts. Thus, it is not uncommon that multiple data acquisitions are performed so that each data set, corresponding to a respiratory phase, has a sufficient number of images for reconstruction. In this work we evaluate the quality of 4D CBCT reconstructions computed using the filtered back-projection method of Feldkamp-Davis-Kress (FDK), and simultaneous algebraic reconstruction (SART), using a simulation framework that mimics the current clinical CBCT acquisition protocol for soft tissue imaging. Methods To evaluate the effect of reconstruction algorithm on the quality of 4D CBCT data we performed the following simulation study. We use a digital phantom composed of two sphere's. The abdomen is represented by a stationary sphere and a target tumor is represented by a sphere performing sinusoidal modulated cranial-caudal motion with an amplitude of 20mm. Amplitude magnitude is based on published studies describing respiratory motion of the liver. Simulated breathing frequencies were set to 15 and 18 breaths per minute, normal respiratory rates. We then generate Digitally Reconstructed Radiographs (DRR) using the projection parameters for soft tissue imaging from a clinical system, the Siemens Axiom Artis dFA. The resulting data set consists of 543 DRRs spanning 220 degrees, with an acquisition time of 20sec. For each of the breathing frequencies we evaluated the quality of the reconstruction algorithms using five and ten subsets per respiratory cycle, with projections binned according to their respiratory phase. Figure 1 shows the simulation setup with five bins and a respiratory rate of 15 breaths per minute. Reconstructed images from each subset were compared to the known scene at that respiratory phase using normalized cross correlation (NCC), Pearson's r. That is, two images are equivalent if they follow a linear relationship. This reflects the equivalence between images mapped using the common radiologic window and level linear mapping. Finally, we also compared image quality of the SART method with an initial reconstruction obtained using the FDK method. Results In all reconstructions the SART algorithm exhibited a slightly higher correlation than the FDK method. When using five bins and a respiratory rate of 15 breaths per minute the FDK method had a NCC value of 0.71 and the SART had a value of 0.76. This setup was the most favorable, in that the subset sizes were the largest in our simulation study. When we increased the number of bins to ten the FDK method had an NCC value of 0.68 and the SART a value of 0.78, reflecting the sensitivity of FDK to the number of projection images. When we used five bins and a respiratory rate of 18 breaths per minute the FDK method had a NCC value of 0.73 and the SART a value of 0.75. When we increased the number of bins to ten the FDK method had a NCC value of 0.7 and SART had a value of 0.78. For the SART algorithm initialized with the FDK result we obtained similar values to SART, only with a reduced computational time. Figure 2 shows the results we obtained for reconstructions using X bins and Y breaths per minute: (a) ground truth; (b)FDK; (c) SART; and (d) FDK followed by SART. For the FDK algorithm reconstructions took on average 1-2 minutes using 100 images. For the same data sets the SART algorithm took approximately 30 minutes. Finally, the combination of FDK followed by SART took approximately 15 minutes. It should be noted that we implemented the SART and FDK algorithms in a straightforward manner and did not optimize our code, as a result the running times of our reconstructions are not optimal. Conclusion Based on our simulations, we see that with the current clinical imaging protocol the SART method is able to produce better reconstructions. This is primarily due to the small number of unique images included in each subset. The fast acquisition time of 20sec with normal respiratory rates results in subsets that are acquired from a limited number of angles. While the number of images included in each subset is substantial they are concentrated around a limited number of viewing angles. Based on these observations we will further investigate the use of the SART method to perform our reconstructions. We will focus on improving the implementation efficiency. This is a promising avenue as published works have shown that using hardware acceleration, with the graphics processing unit, SART reconstructions can be performed in near real time.     «

Introduction C-arm based Cone-Beam CT (CBCT) imaging enables the in-situ acquisition of three dimensional images. In the context of image-guided interventions this technology has the potential to reduce the complexity of a procedure's workflow, with imaging and intervention carried out in the same location. For interventions carried out in the thoracic-abdominal regions, image acquisition while the patient is freely respiring results in blurred images. To improve image quality, data is...     »