Online program Overview Session: MIC-13

Abstract

Fast timing with Cherenkov luminescence has attracted significant interest in the medical imaging community over the last decade. While detector time resolutions of 100–200ps FWHM have been reported, we recently demonstrated that current timing estimators fail to fully exploit the prompt nature of Cherenkov emission. In particular, simulations of a small BGO crystal coupled to a state-of-the-art silicon photomultiplier (SiPM) yield a detector time resolution (standard deviation) of σ_t = 158ps when using only the first detected photon, whereas the Cramér–Rao lower bound (CRLB) predicts a ≈ 4.5-fold improvement to σ_t = 35ps for the same configuration.
To approach this limit, we apply a maximum-likelihood estimation (MLE) framework incorporating the full photon arrival time distribution. Using all detected photons, we obtain a harmonic average detector time resolution of σ_t = 35 ± 2ps, demonstrating statistically efficient timing estimation consistent with the CRLB. The MLE framework also enables event classification, identifying variations in Cherenkov photon yield and time biases from light transport and photon production. Simulations suggest a 2-fold timing improvement compared to state-of-the-art classifiers such as the signal rise time.
To validate this approach experimentally, we apply the method to a semi-monolithic BGO photon-counting detector (42 × 2 × 20 mm³) measured in coincidence with a reference detector and read out by 16 SiPMs. Events with different Cherenkov photon yield can be identified, improving time resolution from 199ps (all events) to 165ps and 140ps FWHM for tighter log-likelihood selections. We observe a stronger improvement over the first
timestamp for Cherenkov-rich events and shallow interaction depths.
These results demonstrate that photon-counting-based MLE can fully exploit the information content of Cherenkov emission, enabling a pathway toward CRLB-limited timing performance for improved TOF-PET imaging.

Abstract

Signal readout strategies of positron emission tomography detectors have an inherent compromise between resolution and complexity, where finer photosensor sampling improves interaction localization but increases the number of readout channels and associated hardware demands. Channel-reduction methods are therefore central to detector design, though they typically can incur some loss in performance. Recent advances in deep learning offer a pathway to mitigate this compromise by recovering information lost through reduced readout. In this work, we propose a neural network–based sparse-to-dense mapping that reconstructs high-resolution (8×8) silicon photomultiplier signals from low-resolution (4×4) inputs. Data was acquired using a semi-monolithic scintillator array with pinhole collimation across both the monolithic and depth-of-interaction directions. An initial upsampling network is trained to transform 4×4 signals into 8×8 representations, which are then used as input to a second network for interaction position estimation. Performance is evaluated by comparing positioning accuracy across four configurations: native 8×8 readout, reduced 4×4 readout, upsampled 4×4-to-8×8 signals, and a conventional row–column summation (8+8) scheme. Results demonstrate that the proposed upsampling approach effectively reconstructs higher-resolution signal structure, helping to improve the monolithic and DOI resolution compared to both 4×4 and 8+8 readouts. This approach could therefore remove the need for additional hardware coming with conventional multiplexing techniques, while recovering the benefits of denser photosensor sampling.

Abstract

Last year, we developed diode-based row-column multiplexing boards that can reduce the number of silicon photomultiplier (SiPM) channels by a factor of four without compromising the coincidence timing resolution (CTR). This technology is key to reducing the production costs of next-generation clinical whole-body time-of-flight positron emission tomography (TOF-PET) systems. However, we used one-to-one coupling detectors with a crystal pitch of 3.2 mm, which cannot provide depth-of-interaction (DOI) information. For brain PET applications, for example, DOI measurement capability with a finer crystal pitch is required. On the other hand, our group has developed crosshair light-sharing (CLS) detectors that can offer both TOF and DOI information with a crystal pitch half that of the SiPM pitch. In this work, therefore, we explored the performance of the CLS detectors with the multiplexing board in combination with a TOFPET2 ASIC module for high-resolution TOF-DOI brain PET imaging. The CLS detectors had a 14×14 array of fast-LGSO crystals (crystal pitch=1.6 mm, crystal thickness=15 mm) and an 8×8 SiPM array (Hamamatsu, S14161-3050HS). The multiplexed SiPM signals were digitized by the TOFPET2 ASIC module. We evaluated CTR at eight different detector positions of the multiplexing boards by shifting the CLS detector pairs and a ²²Na source. The diode multiplexing did not degrade the CTR and energy resolution compared to the case without multiplexing. The average CTR of 366±6 ps was obtained across the eight detector positions. In the future, we plan to further improve the CTR by using DOI information.

Abstract

Positron emission tomography can benefit from time-of-flight information between a coincident pair of PET detectors to enhance image quality. Improved coincidence time resolution (CTR) can not only improve image quality but also obviate the need for image reconstruction when the spatial uncertainty derived from CTR becomes sufficiently small to directly localize positron annihilation events within the patient body, potentially making PET systems more compact and flexible. The first reconstruction-free positron emission imaging termed direct positron emission imaging (dPEI), was demonstrated using detectors with 32 ps CTR; however, limitations such as low detection efficiency, lack of energy information due to the use of a pure Cherenkov radiator, and a small active area resulting from the single-channel detector system remain. To address these limitations and make dPEI more practical, we developed the first prototypes of a multi-anode bismuth germanate-integrated microchannel plate photomultiplier tube (MA-BGO-IMP). The MA-BGO-IMP is position-sensitive and has a two-dimensional 4 × 4 anode array with a 5.75 mm readout pitch. At the detector level, the transit time spread and quantum efficiency were measured to be ~40 ps full width at half maximum (FWHM) without circuit optimization and 31.4% at 380 nm without Fresnel reflection correction, respectively. A coincidence experiment was conducted using a ²²Na point source paired with a barium fluoride-integrated microchannel plate photomultiplier tube. The preliminary energy resolution and CTR were 26% FWHM at 511 keV and 56.2 ps FWHM, respectively. These results were obtained using signals from only one anode. At the conference, we will present more detailed characterization of these newly developed MA-BGO-IMPs with 511 keV-based coincidence results, including energy resolution, CTR, and spatial resolution using a pair of MA-BGO-IMPs.

Abstract

Low-dose imaging and small-animal positron emission tomography (PET) require detectors with both high sensitivity and high spatial resolution. In this study, we investigate the performance of high-density scintillator matrices, such as BGO and LYSO:Ce, combined with multi-channel, low-noise, low-power, high-frequency (HF) readout electronics to enhance the Time-of-Flight (TOF) and Depth-of-Interaction (DOI) performance of pixelated detectors. High-frequency SiPM readout has demonstrated significant improvements in coincidence time resolution (CTR) by enabling the detection of ultrafast optical-photon production processes, such as Cherenkov emission in BGO and fast scintillation in LYSO:Ce. Two DOI-encoding detector designs are compared in this work: a double-sided readout and a single-sided readout employing a light guide to achieve DOI encoding through light sharing in laterally depolished crystal matrices. Experimental results demonstrate that they can achieve a timing resolution comparable to that of a standard pixelated matrix while also providing accurate DOI information. In particular, using 20 mm-long LYSO:Ce crystals, the light-sharing approach achieves a CTR of 133 ±  2 ps FWHM together with a DOI resolution of 2.2 ± 0.1 mm FWHM. Using a double-sided readout further improves the performance, reaching a CTR of 90 ± 2 ps FWHM and a DOI resolution of 1.9 ± 0.1 mm. Both DOI encoding methods provide similar DOI performance with LYSO:Ce matrices, while the advantage of collecting light from both ends of the scintillator becomes more evident with BGO crystals. In this case, the double-sided readout achieves a DOI resolution of 4.4 ± 0.1 mm, compared with 7.2 ± 0.2 mm obtained using the light-sharing approach. Future work will investigate the scalability of these detector concepts by extending the study from 16 to 64-channel prototypes and by comparing different readout electronics.