Online program Overview Session: M-LT-03

Abstract

Spatial normalization enables standardized volume-of-interest quantification in brain perfusion single-photon emission computed tomography, but accurate alignment is difficult because of low resolution, noise, partial volume effects, and uptake-dependent variation. We propose a MRI-informed, SPECT-only framework that distills anatomical deformation fields and feature guidance from a magnetic resonance imaging registration teacher. The student disentangles anatomical and uptake-related representations to suppress functional-uptake interference while requiring only single-photon emission computed tomography at inference. In 173 paired imaging datasets, the method improved standardized uptake value ratio consistency between subject and template spaces, especially for fine-grained volumes of interest, supporting MRI-free, single-pass template-based quantification.

Abstract

The technology of amorphous silicon-based micro-channel plates (AMCPs) uses a versatile approach to stack hydrogenated amorphous silicon in thicknesses up to 100 μm, using plasma-enhanced chemical vapor deposition  (PECVD) and etch microchannels of diameter below 2 μm every 5 μm, on a hexagonal pattern, by deep reactive ion etching. With gains up to 1000 for reverse bias less than 300 V, the design is possible to build on arrays of sensitive transimpedance amplifiers, integrated in read-out circuits with high channel density, such as for instance 15 by 15 μm. This ensures single photon detection with quantum efficiency up to 30% and noise rates below 50 Hz/cm², if combined with state of the art photocathodes such as the hi-QE from Photonis. Current fabrication runs focus on direct integration of the stack on ASICs, to be integrated with the photocathode over the next months. Measurements show timing response with jitter below 10 ps, making this high-precision photodetector ideal for radiation detectors combining ultra-fast photon emission and low light yield, such as Cherenkov.

Abstract

Diamond-based dosimeters are popular for medical dosimetry in view of their fast response, radiation hardness and tissue equivalence. However, some design improvements to achieve better accuracy in signal readout and approach tissue equivalence are possible by the removal of the metallic contacts, sometimes covering the active area of said devices.
In this contribution, we explore the use of a simple processing and scalable solution to obtain non-metal, low Z electrical contact. Thus, the fabrication of diamond detectors with graphitic contacts at the Microelectronics Institute of Barcelona are shown. Several approaches are explored, some of which never utilised before for radiation detector fabrication, namely pyrolytic photoresist film, nickel catalysis and ion surface graphitisation.
A prototype of such device using the pyrolytic photoresist film process for the first time has been produced and was tested under a 60 kV X-ray tube. We report the fabrication and preliminary performance results of such device. Although the substrate material quality is not ideal for radiation detection, it was possible to observe radiation response using the PPF as an interface material. The prototype features a linearity better than 3% as tested in a laboratory X-ray chamber, proving it a promising technology for detector operation. Finally, the technology challenges are finally discussed to assess the feasibility of PPF-diamond technology for dosimeter batch production.

Abstract

Lutetium oxide–based scintillators are of growing interest for positron emission tomography (PET) because of their high density and high effective atomic number, which provide efficient stopping of 511 keV annihilation photons. Lanthanum-doped (Lu,Y)₂O₃ (LYO:La) is a transparent ceramic variant that retains the favorable stopping power of Lu₂O₃ while being easier to fabricate, making it an attractive candidate for next-generation PET detector arrays. In this work we report a scintillation characterization of LYO:La pixels in two geometries relevant to PET, 1×1×10 mm³ for high-resolution preclinical and dedicated-organ imaging, and 3×3×10 mm³ for whole-body clinical systems. Pixels were cut from LYO:La ceramic disks, polished on all faces, and evaluated with a range of side reflectors (Teflon, Tyvek, ESR film, and Spectralon) in both flat and upright coupling orientations. We measured X-ray-excited radioluminescence emission spectra, ²²Na (511 keV) pulse height spectra, relative light yield, energy resolution, scintillation decay time profiles, and timing resolution. The results illustrate the influence of pixel geometry, coupling face, and reflector choice on light collection and energy resolution, and show that LYO:La pixels offer a promising combination of stopping power, light yield, and timing performance for PET detector modules.

Abstract

Dynamic medical X-ray imaging requires detectors that combine high quantum efficiency, low residual lag, and high spatial resolution without the optical spread and afterglow of scintillator-based flat panels. We report a fully inorganic perovskite-based direct-conversion perovskite X-ray imager fabricated from phase-pure orthorhombic Cesium Lead Bromide (CPB) sensor material and integrated with active-pixel readout arrays. Phase-pure CPB was obtained by controlled crystal growth and converted into dense blade-coated films; non-phase-pure starting materials containing secondary phases produced nonuniform films and degraded images, confirming the importance of phase control for reproducible detector performance. The optimized CPB detector demonstrated high linear X-ray response, sensitivity of 55,400 μC Gyair⁻¹ cm⁻² at 30 keV and 0.5 V/μm, a limit of detection of 417 pGyair s⁻¹, a mobility-lifetime product of 8.0 × 10⁻⁴ cm² V⁻¹, and a calculated detective quantum efficiency near 85% at practical dose levels. In flat-panel imaging, CPB imagers reached 6.9 lp/mm at 20% modulation transfer function on a 100 μm-pitch amorphous-silicon array and resolved 20 lp/mm after integration with a CMOS readout. Most importantly for low-lag medical applications, improvements in phase purity and charge-transport stability reduced residual detector lag from 8.9% to 1.1%. These results show that all-inorganic perovskite direct conversion can provide a path to low-dose, low-lag, high-resolution imaging for fluoroscopy, tomosynthesis, dental cone-beam computed tomography, and other image-guided procedures.

Abstract

Total-body positron emission tomography (PET) increases system sensitivity and enables whole-body dynamic imaging, but its extended axial field of view makes accurate depth-of-interaction (DOI) calibration essential for maintaining uniform spatial resolution and reducing parallax error. Conventional DOI calibration requires mechanically scanned collimated sources or detector-specific reference measurements, which become impractical as the number of detector modules increases. PhysiCal-PET introduces a physics-based neural-network framework for autonomous in-system DOI calibration of PET scanners. The method trains a network to map raw detector signals to DOI using coincidence data acquired with an uncollimated positron source at known positions. For each event, the loss function enforces annihilation geometry by requiring the predicted line of response to be consistent with the source position, thereby providing supervision without measured DOI labels. The framework was tested on realistic PET simulations and applied to data from a real preclinical PET scanner. Using this procedure, the DOI response of all 144 detector modules was calibrated in less than 2 hours of point-source measurement time. These results demonstrate that the annihilation physics can replace labor-intensive DOI calibration hardware and enable rapid, routine calibration of high-resolution PET systems. PhysiCal-PET therefore provides a scalable path toward autonomous calibration for total-body PET and other systems that rely on accurate DOI positioning.

Abstract

A spatio-temporal denoising training strategy for low-dose fluoroscopy is proposed, based on a spatially-adaptive temporal consistency metric. This approach acts on a small set of consecutive frames coupling a conventional denoising operator with an adaptive motion-gradient attention gate to selectively modulate the temporal consistency terms within the loss function, ensuring effective noise suppression and flickering reduction while maintaining high anatomical fidelity in dynamic regions.
The proposed method was used to train a denoising operator in which a three-frame temporal window is used as input. The model was trained across multiple dose reduction factors (qF), and demonstrated superior structural recovery and enhanced temporal stability in the processed image sequences, compared to single-frame models. Quantitative and qualitative evaluation of image quality, along with stability maps—defined as the Normalized Cross-Correlation (NCC) of temporal Coefficient of Variation (CV) maps—confirms high fidelity relative to high-dose reference acquisitions, effectively mitigating stochastic fluctuations without introducing motion artifacts.

Abstract

Carbon-ion radiotherapy (CIRT) combines high biological effectiveness with outstanding treatment precision. Since treatment plans are based on the most recent Computed Tomography (CT), anatomical changes during multi-week treatments may compromise dose delivery accuracy. To address the increasing uncertainty about the patient’s anatomical state with time elapsed since the planning CT, our group developed an in vivo monitoring method using secondary charged particles emitted during therapy. This non-invasive approach requires no additional dose and reduces uncertainty about patient anatomy and thus, loss of treatment accuracy. It uses a custom detection system to detect nuclear fragments emmited from the patient and allows approximating their origin. To reconstruct these changes in 3D, we utilize our recently published method for detecting the depth (z) of changes along the beam axis and apply it to a sliding window (x,y) along the transverse plane. However, splitting reconstructed nuclear fragment origins into smaller subsets reduces the statistical stability of each subset distribution. In this study, we present an approach to robustly distinguish whether a windowed subset belongs to a region with or without an anatomical change based on correlation matrices generated by our depth detection method. These images contain location-wise correlations of joined frequency band variations between projections of nuclear fragment measurements from different treatment days. To classify regions of change and approximate change depth, we train an ensemble of deep learning-based image classifiers with different architectures (i.e. ResNet, RegNet and ConvNeXt) and combine their outputs into an ensemble decision. The approach was validated using dedicated irradiation measurements with coin-sized cavities inserted into PMMA head phantoms. Despite limited training data, we achieved robust improvements in change localization using an ensemble of different deep learning architectures.

Abstract

Diffusion models have shown strong potential for improving PET image quality. However, existing methods based on unconditional pretrained diffusion priors often struggle to balance realism and fidelity, particularly in preserving high-resolution details. To address this issue, we propose MI-ResShift, a zero-shot mutual information-guided diffusion model with residual shifting for PET image super-resolution. MI-ResShift introduces a differentiable local-wise mutual information layer to guide the denoising process using structural information from CT or MRI, enabling robust cross-modal feature alignment in the statistical domain. By combining mutual information guidance with residual shifting, the proposed method enhances PET detail fidelity and robustness while reducing diffusion-based inference to only four steps. Experiments on PET/MR brain datasets demonstrate that MI-ResShift improves image quality over classical interpolation methods and provides an efficient solution for cross-modal PET super-resolution.

Abstract

Neural Network (NN) models have emerged as a promising tool for determining 3D photon impact positions in Positron Emission Tomography (PET) detectors based on monolithic scintillators. Nevertheless, NNs typically require long acquisitions of reference data training, becoming a time-consuming process that requires extensive experimental data for each individual detector module. To reduce calibration time, we propose a new methodology based on NNs trained entirely using simulated data to predict the 3D impact position of gamma rays. The feasibility of the proposed approach was verified using a real preclinical PET system composed of two rings of 8 LYSO:Ce monolithic scintillators each.

The simulation of a single detector was implemented in GATE and reference data moving a pencil beam in 1 mm steps was simulated to train two Multilayer Perceptrons (MLPs), MLP_X and MLP_Y, to predict the x- and y- impact positions, respectively. To validate the introduced methodology at reconstructed image level, experimental data was acquired using a ²²Na point source at the center of the preclinical system and processed using both the presented framework and the system’s standard calibration methodology.

Results have demonstrated that similar spatial resolution was obtained with both approaches, thus demonstrating that NNs trained exclusively with Monte Carlo simulations can generate coherent reconstructed images and thus eliminate the need for calibration measurements for each individual detector module.