Neural network turns a thin scintillator into a sharp low-energy gamma-ray camera

Scintillation detectors underpin medical imaging, industrial inspection and homeland security, and they also serve high-precision gamma-ray spectroscopy at facilities such as ELI-NP. When spatial information matters, engineers have traditionally segmented the scintillator into many small crystals, each with its own readout channel. That works, but segmentation multiplies manufacturing complexity and introduces dead space between crystals, degrading both spectroscopic performance and detection efficiency. The alternative is a monolithic crystal watched by a segmented readout, where the position of each interaction must be inferred from how the scintillation light is distributed across neighboring pixels.

State-of-the-art monolithic detectors have already achieved sub-millimeter resolution for photons in the megaelectronvolt range, but the published literature concentrates on energies at or above 511 kiloelectronvolts. At 59.5 kiloelectronvolts, the energy of the gamma rays emitted by the decay of americium-241 that the team chose as their test case, each interaction produces far fewer optical photons, and the signal-to-noise ratio deteriorates accordingly. The authors note that using 59.5 kiloelectronvolt gamma rays results in a lower signal-to-noise ratio, owing to an eightfold reduction in optical photons compared with higher-energy conditions, and that their methods achieve comparable performance in these more challenging conditions.

The 70 by 70 millimeter crystal is coated with reflective epoxy on all but one face and coupled with optical grease to an Onsemi Array-J series silicon photomultiplier array whose cells measure 6.13 by 6.13 millimeters on a 6.33 millimeter pitch, giving a total active area of about 50 by 50 millimeters. A CAEN DT5202 module processes each of the 64 channels individually in spectroscopy mode, so every interaction event is recorded as a timestamp plus a complete 8 by 8 image of signal amplitudes. The crystal was deliberately made larger than the sensor array; an earlier version built around a 50 by 50 millimeter crystal had produced three distinct classes of spectra, at corners, edges and inner cells, because of edge reflections, and the alignment algorithm could not reconcile them. Enlarging the crystal eliminated those differences.

Aligning the energy response of 64 separate channels proved to be one of the study’s central technical hurdles. Normally, calibration relies on finding a photopeak in each channel’s spectrum, but the small cell size and the low number of emitted photons mean no such peak exists at the individual sensor level. The team combined hardware adjustments, tuning applied voltage and gain channel by channel, with a software stage in which a genetic algorithm searched for linear alignment coefficients. The algorithm used the Pearson correlation coefficient to judge how closely each channel’s histogram matched a reference, running for 20 generations with a population of 20, and bounded the search space for intercepts and slopes.

To train and test the position-recovery models, the researchers scanned a collimated americium-241 source, shaped into a 1.2 millimeter diameter pencil beam by a 20 millimeter thick densimet block, across 1000 randomly distributed points on the detector face, spending five minutes at each point at a counting rate near 4.6 kilohertz. After filtering events to the 59.5 kiloelectronvolt line and rejecting points where the collimator aperture fell partly outside the active region, 974 measurement points remained, split into training, testing and validation subsets in a 70, 15 and 15 percent ratio, with each point contributing roughly 4000 events. Dihedral transformations generated augmented copies of the training and testing data to enlarge the effective dataset.

The first model is analytical: it fits a two-dimensional isotropic Gaussian on a flat baseline, a five-parameter curve, to each 8 by 8 event image. Fitting so few parameters to so coarse an image is unstable and highly sensitive to initialization, so the team bounded four of the five parameters and used a random search to find amplitude and sigma limits that minimized the Euclidean distance between predicted and true positions on the testing data. The second model is a reduced ResNet convolutional neural network with six residual blocks, trained with the Adam optimizer for 250 epochs at a learning rate of 5 times 10 to the minus 4, on an Nvidia RTX A4500 GPU. Both models were then evaluated on the same validation set.

Measuring positional accuracy as the mean Euclidean distance between each validation point’s reconstructed centroid and the true collimator position, the network achieved 0.64 plus or minus 0.05 millimeters, an 11 percent improvement over the analytical model’s 0.72 plus or minus 0.06 millimeters. Precision, quantified as the full width at half maximum of the summed x and y position distributions, favored the network even more strongly: 1.68 plus or minus 0.05 millimeters against 2.21 plus or minus 0.02 millimeters, a 24.1 percent advantage. Both figures were estimated with bootstrap resampling. Discarding edge points improved accuracy for both models but narrowed the relative gap to about 9 percent, and precision degraded for both, a sign that neither model handles the array border well.

The imaging test made those strengths and weaknesses visible. With an uncollimated source 22.5 centimeters from the detector and a 5 millimeter stainless steel mask in front, the raw pixel-center image barely separated the mask’s features, underscoring the hard limit of an 8 by 8 readout. Both models recovered far more detail, though each imprinted its own artifacts: the analytical model produced a symmetric grid of hot spots at cell centers, while the network’s pattern was more irregular, and in places its artifacts could be mistaken for real mask features. After dividing by a flat-field image, collected over 17 hours without the mask, and applying a 15 percent intensity threshold, most artifacts vanished and every mask feature became distinctly visible in both images, though distortions lingered along the mask edges.

The authors note their accuracy and precision figures are competitive with a reported resolution of 1.78 millimeters for a similar readout array, achieved here despite the lower light yield of 59.5 kiloelectronvolt photons and the thin 2 millimeter crystal, which restricts light spread and amplifies pixel-center artifacts. The analytical model converged to a consistent solution faster and produced more visually uniform images, and its fitted Gaussian width correlated with the expected exponential depth distribution, hinting at weak depth labels for future training. The network, which makes no single-interaction assumption, is expected to be faster at inference, and a larger development dataset should lift its performance further. For now, the demonstration stands: a thin monolithic scintillator, read out by a modest sensor grid, can image low-energy gamma rays at a resolution that significantly surpasses the readout array itself.

Subject of Research: Position-sensitive gamma-ray detection with a thin monolithic scintillator and neural network position reconstruction at low energies

Article Title: A position-sensitive scintillation detector for low-energy gamma-rays

Article References: Turturica, G. V., Temelie, B., & Iancu, V. (2026). A position-sensitive scintillation detector for low-energy gamma-rays. EPJ Techniques and Instrumentation, 13(1), Article 1. https://doi.org/10.1140/epjti/s40485-026-00120-2

Image Credits: AI Generated

DOI: 10.1140/epjti/s40485-026-00120-2

Keywords: gamma-ray imaging, scintillation detector, position-sensitive detector, silicon photomultiplier, neural network, CsI(Tl), monolithic scintillator, sub-pixel resolution, americium-241, energy alignment, ResNet, medical imaging