Online program Overview Session: JS-02

Abstract

Photon-counting computed tomography (PCCT) requires scintillators with ultrafast decay times, ultralow afterglow, and high light yields (LY) to prevent pulse pile-up and image ghosting under continuous high-flux X-rays. While cerium-doped lutetium yttrium oxyorthosilicate (LYSO:Ce) is a benchmark material, its intrinsic ~40 ns decay time and deep-trap-induced afterglow constrain its PCCT applications. This work demonstrates systematic defect and crystal-field engineering to modify LYSO:Ce crystals, establishing a highly customizable material platform.

Abstract

Scintillation based photon counting detectors (PCDs) attracts great interest because of its fast time response, crystal separation to suppress the charge sharing and wider choices of sophisticated scintillation crytals. We have designed and fabricated a custom fine-pitch (250 μ m) 10 × 10 Ce:YAGG- SiPM array whose size is comparable to the current commercial compound semiconductor based PCD. The performance was characterized in terms of energy resolution, spatial resolution, count rate capability and imaging capability. The measured results shows 28% @ 122 keV moderate energy resolution, 2.9 LP/mm spatial resolution comparable to the existing CT, 20 Mcps/mm2 at 20 % loss count rate capability. The comparison between YAGG and GFAG shows the significant supression of K-α escape cross-talk in the energy spectrum. The investigation with other scintillation crystals, such as LYSO will be shown. The different concentration of Optiray contrast agent (Iodine) with 10 m-50 mg/mL was clearly identified with 4 thresholds photon counting system.

Abstract

In this work, the latest developments in glass-less Through-Silicon Via (TSV) integration of FBK NUV-HD-MT Silicon Photomultipliers (SiPMs), targeting next-generation Time-of-Flight Positron Emission Tomography (ToF-PET) detectors in the framework of the PetVision project, are presented. The NUV-HD-MT technology platform provides an outstanding baseline for TSV integration, featuring Photon Detection Efficiency (PDE) exceeding $65\%$ at $420\ \text{nm}$, optical crosstalk below $5\%$ at $15\ \text{V}$ excess bias, and a Single-Photon Time Resolution (SPTR) as low as $\simeq 18\ \text{ps}$ FWHM for a single SPAD. The glass-less TSV architecture eliminates any permanent intermediate layer at the scintillator interface by removing the temporary glass carrier after backside thinning to $150\ \mu\text{m}$, enabling optimal optical coupling and preserving the fast-timing performance of the underlying technology. The TSV development follows a structured, learning-cycle-based R\&D approach. Preliminary characterization of the Learning Cycle 1 (LC1) front-side production confirms that timing performance remains at state-of-the-art levels, with a Coincidence Time Resolution (CTR) of $\simeq 75\ \text{ps}$ FWHM achieved using a $3\times3\ \text{mm}^2$ SiPM coupled to a $2.8\times2.8\times3\ \text{mm}^3$ L(Y)SO:Ce:Ca scintillator, and an SPTR of $\simeq 25\ \text{ps}$ FWHM obtained for a single $50\ \mu\text{m}$ SPAD. Back-side processing of the LC1 wafers has been completed and full electrical and functional characterization is now underway. The Learning Cycle 2 (LC2) design, featuring an optimized $4\times2$ SiPM array with a $3\ \text{mm}$ channel pitch for direct crystal matrix coupling and a reduced TSV count of one per channel for improved yield, represents a significant step toward a robust, production-ready solution.

Abstract

In time-of-flight PET, coincidence time resolution (CTR) is essential, with the ultimate aim of reaching 10 ps CTR [1]. Cherenkov radiation in BGO crystals, offers a path toward this target, but its low photon yield demands detectors combining high photon detection efficiency, low dark count rate, and fast single-photon time resolution (SPTR) [2].

In DIGILOG, we segment state-of-the-art analog SiPMs into smaller clusters called µSiPMs [3]. With fewer SPADs per cell, µSiPMs have lower capacitive loading, enabling simpler, faster electronic interfaces and higher-resolution timestamping via time-to-digital converters (TDCs) in proximity through 3D stacking. Optical and system-level simulations identify a granularity of 6×6 SPADs per µSiPM (100 µSiPMs in 3×3 mm²) as optimal for Cherenkov-based BGO timing [4].

The µSiPM top tier, fabricated by FBK in NUV-HD-MT technology [5], uses Single-Cell TSVs for independent microcell cathode access without fill-factor loss. We present two 110 nm CMOS readout architectures, both based on a current-mirror front-end, a Vernier ring-oscillator TDC [6], and a serial interface. Two ring oscillators run at slightly different frequencies setting the LSB; the sequences are compared as in a Vernier line until pulses align, and T is computed from the counts n1 and n2. The first implementation, an 8×8 TDC array with 20 ps target LSB, has been fabricated; a Vernier LSB of 19 ps and 4.4 ps RMS jitter have been measured [7]. The second upscales to a 9×9 array with three TDCs per µSiPM on independent threshold levels of a multi-level comparator, enabling MD-SiPM-like operation [8] and order-statistics-based timing [9]. It targets a 6 ps LSB at 2.0 mW per channel, with an on-chip PLL (<5 ps FWHM jitter) stabilizing the oscillators against PVT. Post-layout simulations yield SPTR <20 ps (FWHM) and INL ±1.1 LSB. Die-to-die 3D bonding of the two tiers is ongoing at T-Micro.

Abstract

Depth-of-interaction (DOI) information is critical for improving time-of-flight (TOF) precision in positron emission tomography (PET) detectors. Traditional DOI estimation methods, such as dual-ended readout and layered crystal arrays, require complex fabrication or additional photodetector channels, increasing system cost and degrading timing performance. This paper presents a machine learning framework for real-time DOI layer prediction in single-ended readout LYSO crystal arrays (5×6 structure coupled to a 2×2 SiPM array). A neural network takes normalized SiPM energy ratios and crystal position as inputs and outputs DOI classification into 2, 3, or 6 depth layers. A symmetry-based training strategy is introduced to enhance model generalization across crystal positions, effectively quadrupling the training data through rotational coordinate transformations. Experiments with a collimated ²²Na source demonstrate classification accuracy of 40–80% across six depth layers. Validation using top radiation data confirms the model's generalization from lateral irradiation (training) to normal incidence (clinical deployment). DOI-corrected coincidence time resolution (CTR) evaluated on 4320 crystals shows improvements of 2.25 ps (2-layer), 3.41 ps (3-layer), and 5.18 ps (6-layer).  The two-layer neural network has been implemented on FPGA with fixed-point arithmetic, achieving a processing rate  for real-time inference. Ongoing work focuses on system-level integration and image quality evaluation.

Abstract

Time-of-Flight Computed Tomography (ToF-CT) enhances contrast-to-noise ratio by suppressing patient-scattered photons. However, the fine pixel pitch required for sub-millimeter spatial resolution intensifies inter-crystal scattering (ICS), a confounder that conventional energy-based discrimination cannot resolve under polychromatic X-ray sources, ultimately compromising spatial resolution. To overcome this limitation, we introduce the first real-time, physics-informed ICS correction framework deployed at the FPGA-based detector readout of ToF-CT systems. Built on Physics-Informed Knowledge Distillation (PI-KD), the framework compresses a high-capacity teacher (AUPRC = 0.99) into a compact student architecture suitable for embedded deployment. Specifically, the distillation objective constrains the student to remain consistent with Klein–Nishina scattering kinematics and Monte Carlo-derived interaction priors, thereby preserving the physical structure of its spatiotemporal and energetic representations under aggressive compression. Quantization-Aware Training (QAT) further enables hardware-native INT8 inference, reducing the parameter footprint to ∼2.3 kB with negligible performance loss. Post-synthesis on a Xilinx ZCU104 at a 200 MHz target clock confirms real-time operation beyond the 1 MHz acquisition target, sustaining 1.32 MHz per inference instance with 0.83 µs latency, while INT8 LUT-based arithmetic leaves all DSP blocks free (28% LUT, 38% BRAM, 0% DSP).

Abstract

We have developed a bismuth germanate (BGO) Cherenkov and Scintillation photon counting time-of-flight positron emission tomography (TOF-PET) detector which can count and provide unique timestamps for detected optical photons. This allows 511 keV photon time of interaction estimators to be derived from the prompt Cherenkov light without influence from scintillation light, which has a much higher variance from time of detection. This detector concept is also capable of three-dimensional positioning of 511 keV photon interactions, allowing for position and depth-of-interaction dependent optical transit time skew corrections to be applied to the prompt estimators. Information from this detector concept can be used to classify each detected optical photon as Cherenkov or scintillation. In this work, we show how labeling each detected photon as Cherenkov/scintillation and only incorporating prompt information into a timing estimator that averages timestamps for events with more than one detected Cherenkov photon significantly improves both achievable coincidence time resolution with BGO and also the TOF kernel shape by significantly reducing the long tails from time pickoff on scintillation light. We constructed a 16x16x15 mm³ BGO photon counting detector prototype and evaluated its performance in coincidence with a fast reference detector. Incorporating the information from multiple labeled Cherenkov photons improved CTR from 159 ps to 137 ps FWHM, and the long scintillation tail was reduced to a small fraction of the total coincidence distribution. When considering all events with at least one Cherenkov photon detected, CTR was 122 ps FWHM with a Gaussian kernel shape. CTR improved to ≤100 ps for events containing 2 or more detected Cherenkov photons. We present these methods, along with a new detector concept with high Cherenkov photon detection efficiency to enable ≤100 ps CTR in a BGO TOF-PET detectors that also have high 511 keV photon detection efficiency.

Abstract

Proton therapy provides high dose conformity, improving protection of organs at risk, especially for radioresistant tumours, sensitive locations, and paediatric cases. However, reported toxicities may be linked to high linear energy transfer (LET) effects. To address this issue, IMB-CNM has developed a new radiation detector architecture [1]: 3D-cylindrical silicon microdetectors, with micrometre-scale sensitivity to quantify LET. Starting from this kernel, different configurations have been designed and fabricated to cover a wide range of resolutions and radiation sensitive areas from a 11 × 11 array (2 mm × 2 mm) [2] to a linear array of 200 of those microdetectors covering  0.4 mm × 12 cm [3, 4], among others. To the best of our knowledge, these are the largest radiation sensitive surfaces covered with microdosimeters so far. Each of these multi‒arrays are integrated in customized electronics readouts by using multichannel ASICs to DAQ. The ASIC enables the beam trigger and an analogue output for 128 microdetectors. Additionally, special pitch‒adapters have been manufactured for each configuration to connect the signal channels with the ASIC readouts properly. 
In this work we present the first measurements in proton therapy centres under clinical fluence rates (10^8 p/cm2s) across different regions of the beam. Additionally,  Geant4-based Monte Carlo simulations were  performed to crosscheck the experimental results. Representative microdosimetry energy spectra were obtained for some clinical depths as well as the comparison with the patient simulated LET values obtained by a treatment planning software.
The results show the feasibility of quantifying distributions of microdosimetry lineal energy values under clinical conditions, which are compared with clinically simulated for the first time.