Comparative Evaluation of Inter-Detector Scatter Recovery Methods for Small Animal PET

Objectives: Inter-detector scatter (IDS) is a major type of triple coincidence events in PET which occurred by the Compton scattering of a photon from a detector block to another. Finding the earliest interacted detectors, or recovering IDS events, is a potential way to enhance sensitivity of PET scanner by utilizing line-of-responses (LORs) which have been practically rejected. The IDS occurs frequently especially in small animal PET scanners because detector blocks neighbored closely. Several methods have been proposed to recover IDS effectively while maintaining good image qualities. In this study, the performances of three promising IDS recovery methods (the maximum energy method, the Compton kinematics method and the neural network method) for small animal PET scanners were compared using Monte Carlo simulation data. For the optimization of IDS recovery methods, various energy criteria were applied.

Methods: In most IDS events, a photoelectric absorption of an unscattered photon occurs in the detector block (P) where the largest energy is deposited (E_P), while scattered photons are detected in the other two (S₁, S₂) with the energy of E_S1 and E_S2. Three recovery methods were applied to find the first interacted detector of S₁ and S₂. First, in the maximum energy method (ME), the detector block with larger energy detected was determined as the first interacted one. The Compton kinematics method (CK) chose the detector which best fits Compton kinematics by using two angles derived by geometric information. Third, in the neural network method (NN), the network trained using deposited energies and scatter angles of simulated IDS events was applied to test datasets to determine the first interactions. All methods were applied to a small animal MR-compatible PET system with the inner diameter of 6 cm and the axial length of 5.5 cm simulated by GATE (v.7.0). Sensitivity increase by adding recovered IDS events to original double coincidence events was investigated according to different energy thresholds. For each method and energy threshold, accuracy of recovery was calculated as the ratio of correctly recovered among total IDS events. NEMA NU-4 IQ phantom was simulated and reconstructed for image quality evaluation. Uniformity, signal-to-noise ratio (SNR), contrast and recovery coefficient (RC) were used as figures of merit to compare images which IDS recovery methods were applied to.

Results: In all three methods, sensitivity increase by IDS recovery was 20-36% depending on energy thresholds (250-450 keV). Larger improvement of sensitivity was observed with a lower threshold since the larger IDS events within broader criteria was used. Therefore, uniformity and SNR of images were much enhanced. On the contrary, accuracy was better in higher energy threshold because large fractions of scattered events, which can give false information of energy or angle, were rejected. The NN method achieved the best accuracy among the methods in overall energy criteria. At 350 keV threshold, accuracy of NN was 70% while those of ME and CK were 59% and 61%, respectively. The simplest method, ME, showed inferior accuracy in 250 keV threshold but it was comparable with CK in some higher thresholds. The SNR, contrast and RC were better when a method of higher accuracy was applied since the number of false LORs was reduced.

Conclusion: All the IDS recovery methods tested in this study showed some significant sensitivity improvement which can achieve shorter scan times or lower dose injections. Most importantly, the best accuracy of neural network method with high energy threshold resulted in the finest image qualities. Considering the tradeoff between sensitivity increase and image qualities, optimizing IDS recovery in PET will be conducted as a future work. Research Support: This work was supported by grants from the National Research Foundation of Korea (NRF) funded by the Korean Ministry of Science, ICT and Future Planning (grant no. NRF-2014M3C7034000 and NRF-2016R1A2B3014645).