Dictionary learning based panel PET image reconstruction

In our previous work, we have proposed a stationary panel positron emission tomography (PET) for human body imaging. Besides the large solid angle, the panel PET also provides an open geometry which allows other operations being performed simultaneously during a PET scanning. However, the main challenge of panel PET is the limited view problem. Images reconstructed using traditional algorithms will suffer serious artifacts. Time-of-flight (TOF) information can be incorporated into the image reconstruction to reduce the artifacts. However, the best system timing resolution available in current commercial PET scanners is about 500 ps, which is not good enough to achieve satisfactory images. In this paper, a novel list mode TOF reconstruction algorithm coupled with dictionary learning was formulated to improve the panel PET image quality. The dictionary is an adaptive sparsifying transform and can be learned from high-quality training images. So images being estimated have a sparse representation with respect to the learned dictionary. With this prior knowledge, the dictionary is used in the reconstruction algorithm to remove artifacts and noise. The results show that, compared with ordered subset expectation maximization (OSEM), reconstruction with adaptive dictionary can recover more details of the object boundaries and substantially enhance the image quality. The results also imply that, dictionary learning may help to relax the requirements on TOF and realize the panel PET tomographic image reconstruction under the current TOF technique.