Room: AAPM ePoster Library
Purpose: To develop and characterize a novel kernel-weighted back-projection (KWBP) algorithm that reduces noise in Compton camera (CC) imaging.
Methods: A GPU optimized KWBP algorithm for image reconstruction from the datasets measured by CCs was developed. In this algorithm, all the half-cones, which are derived from double or triple scattered events and contain the origin positions of gammas on their surfaces, are projected into a voxelated image space. The gamma emission probability in each voxel is calculated using the Epanechnikov kernel density of the minimum distance between the voxel center and the surface of each cone. To suppress the noise in the image, all the cones are randomly shuffled and back-projected to the image space and the resulting image is subtracted from the calculated image. Two key parameters of the KWBP algorithm, namely bin-size and bandwidth that affect the quality of the images and ability to predict the true source location were characterized. Two identical prototype CCs positioned orthogonally to the source were used to measure the 0.662 MeV gammas from a Cesium-137 source at four different locations.
Results: A bandwidth value between 6 and 14 was found to best estimate the true locations of the source. The positions of the source estimated by KWBP reconstructed images were found to be off by 3.0±1.4 mm, 3.8±1.4 mm, 3.8±1.4 mm, and 8.4±1.4 mm from their respective true positions. The location of the fourth source was much further away than the other three locations from the CCs.
Conclusion: Leading CC imaging reconstruction algorithms such as back projection and list-mode maximum-likelihood expectation maximization do not account for noise. The noise suppression ability of KWBP algorithm improves the efficiency of CC imaging and can be used to improve the online range verification of proton therapy.