Total variation and point spread function priors for MLEM reconstruction in Compton camera imaging

Abstract

The Compton camera is a gamma ray imaging device already employed in astronomy and still in investigation for clinical domain. A key point in the imaging process is the tomographic reconstruction step. When the acquisition parameters and the a priori information are correctly accounted for, iterative algorithms are able to produce accurate images by compensating for measurement uncertainties and statistical noise. In this work we focus on the list-mode maximum likelihood expectation maximization (LM-MLEM) algorithm with smoothness a priori information expressed by the total variation norm. This type of regularization is particularly well suited for low-dose acquisitions, as it is the case in the applications foreseen for the camera. We show that the TV a priori strongly improves the images when data are acquired in ideal conditions. For realistic data, this a priori is not sufficient and deconvolution with a pre-calculated image-space kernel should also be considered.

Development and Applications of Compton Camera—A Review

“…However, the convergence speed of the adopted iterative reconstruction method was not sufficiently high because the imaging algorithms considered the high-energy γ-ray source located at a far distance, and the simulation programs developed at that time considered the simple detector system. Some studies have recently been conducted to improve the MLEM and compare it with other imaging models, with a focus on the application of Compton cameras in nuclear medicine and particle therapy [ 81 , 210 , 211 , 212 , 213 ]. Yabu et al [ 214 ] reported the superiority of the LM-MLEM method over SBP using a Si/CdTe Compton camera.…”

Hybrid PET/Compton-camera imaging: an imager for the next generation

“…), the concept of COR can be generalized to a volume of response (VOR), which represents the region in a 3D space where the origin of the event originates. Several approaches have been introduced to model the VORs into the reconstruction, see, for example [83,[88][89][90][91][92]. Among them, including a model of the point-spread-function (PSF) in image space as a convolution kernel is an extended option, see Fig.…”

“…Using specific edge-preserving priors and appropriate β values, image noise can be effectively reduced without loss of resolution. In spite of the great potential of Bayesian reconstruction, only few works report the use of priors for Compton-camera data [105,[108][109][110].…”