Abstract
1514
Introduction: Cadmium zinc telluride (CZT) has a high Compton scatter fraction (relative to the total absorption coefficient) at 511 keV. When using large uniform CZT crystal block (4×4×0.5 cm3) with an edge-on orientation to build positron emission tomography (PET) systems, the intra-crystal scatters and inter-crystal scatters can be detected. As a consequence, there is an opportunity to improve the system sensitivity by considering multiple interaction photon events (MIPEs). The position identification of the first interaction event is of high importance as it contributes to the line of response (LoR) generation for image reconstruction. In this work, we have investigated the application of machine learning to identify the correct interaction sequence in MIPEs in a PET system based on CZT.
Methods: In order to compare the performance of our machine learning approach for correct event sequencing accuracy with an existing method such as Compton kinematics, we used the small animal PET system described in [1]. This system and NEMA NU4 image quality mouse phantom was simulated in GATE. The system energy resolution was reported to be 7.35 ± 1.75% full-width-at-half-maximum (FWHM) at 511 keV and its time resolution is 18 ± 0.1 ns FWHM. This system can achieve 1 mm spatial resolution in all three directions within field-of-view. The recorded events in the output file of GATE simulation (hit file) were blurred and binned based on the system performance. Given the energy threshold of 10 keV, the events with recorded energy below the threshold were discarded. Then, events falling within the time window of 20 ns were considered as coincident events. The energy window was selected between 429-593 keV. If the sum of the energy of detected events in one panel was out of the energy window, the coincidence event was rejected. Note that photoelectric-photoelectric interactions were discarded since the aim of this work is solving the sequence ambiguity for photons that undergo Compton scatter(s) followed by a Photoelectric absorption. All the possible scenarios for the sequence of each MIPE were considered in the data set and for each scenario θP and θE were calculated. θE is the Compton scattering angle (based on energy) and θP is determined as the angle between the incident and scattered photon (based on detected interaction position). Using the ground truth in GATE simulation, the correct sequence and therefore the true scenario for each more-than-two-interaction events is known. This scenario was labeled correct while the rest were labeled wrong. Energy of the first and second interactions as well as θP and θE were selected as features. For the purpose of correct sequence prediction, the support vector machine (SVM) classifier was trained and tested using a leave-one-case-out (LOCO)-based cross validation method.
Results: Sequence identification by employing machine learning results in an improvement in percentage of coincidences with correctly-ordered interaction sequences compared to Compton kinematics method. Compton kinematics recognized 64.9% of two-interaction events correctly [2], while machine learning classifier identified the correct position for the 77.7% of MIPEs. The table below shows the percentage of events for which machine learning predicted the correct orders.
Conclusions: This study demonstrates the application of machine learning to accurately determine the correct order of events and subsequently the first interaction position in MIPEs. In other words, machine learning has potential to be used for inclusion of MIPEs rather than discarding them and improving system sensitivity. References: [1] Abbaszadeh et al 2016 Phys. Med. Biol. 61 6733 [2] Abbaszadeh et al 2018 Phys. Med. Biol. 63 025012