A CNN-based four-layer DOI encoding detector using LYSO and BGO scintillators for small animal PET imaging

<i>Objective</i>. We propose a novel four-layer depth-of-interaction (DOI) encoding phoswich detector using lutetium-yttrium oxyothosilicate (LYSO) and bismuth germanate (BGO) scintillator crystal arrays for high sensitivity and high spatial resolution small animal PET imaging.<i>Approach</i>. The detector was comprised of a stack of four alternating LYSO and BGO scintillator crystal arrays coupled to an 8 × 8 multi-pixel photon counter (MPPC) array and read out by a PETsys TOFPET2 application specific integrated circuit. The four layers from the top (gamma ray entrance) to the bottom (facing the MPPC) consisted of a 24 × 24 array of 0.99 × 0.99 × 6 mm³LYSO crystals, a 24 × 24 array of 0.99 × 0.99 × 6 mm³BGO crystals, a 16 × 16 array of 1.53 × 1.53 × 6 mm³LYSO crystals and a 16 × 16 array of 1.53 × 1.53 × 6 mm³BGO crystals.<i>Main results</i>. Events that occurred in the LYSO and BGO layers were first separated by measuring the pulse energy (integrated charge) and duration (time over threshold (ToT)) from the scintillation pulses. Convolutional neural networks (CNNs) were then used to distinguish between the top and lower LYSO layers and between the upper and bottom BGO layers. Measurements with the prototype detector showed that our proposed method successfully identified events from all four layers. The CNN models achieved a classification accuracy of 91% for distinguishing the two LYSO layers and 81% for distinguishing the two BGO layers. The measured average energy resolution was 13.1% ± 1.7% for the top LYSO layer, 34.0% ± 6.3% for the upper BGO layer, 12.3% ± 1.3% for the lower LYSO layer, and 33.9% ± 6.9% for the bottom BGO layer. The timing resolution between each individual layer (from the top to the bottom) and a single crystal reference detector was 350 ps, 2.8 ns, 328 ps, and 2.1 ns respectively.<i>Significance</i>. In conclusion, the proposed four-layer DOI encoding detector achieved high performance and is an attractive choice for next-generation high sensitivity and high spatial resolution small animal positron emission tomography systems.