Reduction of random coincidences in small animal PET using Artificial Neural Networks

Accidental coincidences (randoms) are one of the main sources of image degradation in PET. In conventional PET, coincidence identification is usually carried out through a coincidence electronic unit, so that randoms occur when two photons arising from two annihilation events are detected within the same coincidence time window. To compensate for randoms, the number of random events contributing to each line-of-response (LOR) should be estimated. On the other hand, some novel systems allow coincidences to be selected post-acquisition in software, or in real time through a digital coincidence engine in a FPGA. These approaches provide the user maximum flexibility in the post-processing of acquired data, thus allowing alternative coincidence sorting procedures to be applied. In this work a natural application of Artificial Neural Network (ANN) methods to the problem of reducing random coincidences has been investigated. It has been compared against a conventional coincidence sorting algorithm based on a time coincidence window combined with geometrical conditions. At matched efficiencies, the ANN based method presents always a sorted output with a smaller random fraction. In addition, two differential trends are found: the conventional method presents a maximum achievable efficiency while the ANN-based one is able to increase the efficiency up to unity, the ideal value, but at the cost of increasing the random fraction. Pattern recognition capabilities of ANN could be more adequate for more complex situations, where simple techniques such as the conventional sorter fails to model the underlying physics.