Online program Overview Session: MIC-19

Abstract

This study proposes a Noise-Aware Brownian Bridge Diffusion Model (NA-BBDM) for low-dose positron emission tomography (PET) denoising. The proposed NA-BBDM models a conditional diffusion path between paired low- and full-dose PET images and incorporates the estimated noise level from the input image into the network. The noise level is estimated from the variance of the residual between the low-dose image and a Gaussian-smoothed version. The diffusion path uses a parabolic variance schedule with zero variance at both endpoints (i.e., low- and full-dose images), parameterized by peak variance and number of diffusion steps. Both are treated as noise-dependent hyperparameters and optimized through grid search. Training data consisted of 25 ¹⁸F-FDG datasets acquired on the PennPET Explorer and subsampled to emulate dose levels from 1/5 to 1/300 of total counts. Performance was compared with a U-Net and evaluated on independent ¹⁸F-FDG and unseen ¹¹C-carfentanil (CFN) datasets. The NA-BBDM produced images with reduced noise speckles (still present with the U-Net) and better-preserved high-uptake regions at ultra-low-dose levels. Quantitatively, the NA-BBDM also achieved lower SUV variability and mean squared error for ¹¹C-CFN, especially below 1/60th dose. Our results suggest that the NA-BBDM provides a robust framework for ultra-low-dose and cross-tracer PET denoising.

Abstract

Low-dose positron emission tomography (PET) imaging can reduce radiation exposure and shorten acquisition time, but reduced photon counts introduce strong count-dependent noise and degrade quantitative reliability. Supervised deep learning methods usually require paired low-dose and normal-dose PET images, which are difficult to acquire in clinical practice. We propose N2N-DiffPET, a self-supervised diffusion framework for low-dose PET denoising without normal-dose supervision. Inspired by recent self-supervised diffusion MRI denoising, N2N-DiffPET adapts iterative diffusion refinement to PET using paired noisy PET observations generated from independent low-dose measurements by event splitting or Poisson resampling. To better model PET noise, N2N-DiffPET constructs a fusion trajectory between the paired noisy PET images and perturbs it using pair-derived PET noise rather than synthetic Gaussian noise. Co-registered MRI is incorporated as an anatomical condition to improve structural preservation. During inference, bidirectional deterministic denoising is performed, and the two directional outputs are averaged to reduce directional bias. Normal-dose PET images are not used for training and are only used for evaluation. Experiments on BrainWeb-based PET/MR simulation data show that N2N-DiffPET suppresses noise while preserving anatomical structures and regional uptake, outperforming conventional denoising methods and self-supervised baselines in visual quality and quantitative metrics.

Abstract

Low-dose PET reduces patient radiation burden by decreasing injected radiotracer dose or scan duration, but the resulting reduction in photon counts introduces statistical noise that degrades diagnostic image quality. Most deep learning denoising methods are trained at a single fixed dose reduction factor and do not generalize to unseen dose levels. We propose Consistency Flow Matching (CFM), a multi-dose denoising framework that augments flow matching with a cross-dose consistency regularization. During training, pairs of low-dose images at two different dose reduction factors are jointly processed, and a consistency loss enforces agreement between their predicted normal-dose outputs, promoting a dose-invariant flow field. Dose conditioning is achieved through Spatial Feature-wise Linear Modulation with
a log-sinusoidal dose embedding. Experiments on the UltraLow-Dose PET Challenge dataset show that CFM improves perceptual image quality and out-of-distribution generalization over standard flow matching, confirming the benefit of cross-dose consistency regularization.

Abstract

Cardiac PET is widely used for myocardial perfusion imaging and viability assessment, both of which require high image quality and quantitative accuracy. In free-breathing acquisitions, respiratory motion can induce cardiac displacement, particularly along the superior-inferior (SI, z-axis) direction, leading to image blurring, attenuation-emission mismatch, and quantitative bias. Reliable estimation of this translation is therefore essential for effective motion correction. In this work, we propose COD-S (Centroid of Distribution-Scaled), a data-driven method to estimate respiratory-induced SI cardiac translation in cardiac PET. COD-S integrates the high sensitivity and high temporal resolution of COD analysis with the physically scaled displacement provided by image registration. Specifically, the respiratory motion trajectory is  decomposed into waveform and amplitude components, where the waveform is derived from COD and the amplitude are calibrated using registration-based estimates. This produces a high-temporal-resolution and physically interpretable cardiac translation trajectory without external respiratory hardware, providing a practical input for motion correction in cardiac PET.

Abstract

High-resolution PET detectors require position dependent energy calibration due to the spatial variability of their optical response. This work addresses that challenge for a novel hexagonal LYSO detector structured by subsurface laser engraving (SSLE) with a double honeycomb displaced crystal array, coupled to a custom 61-channel hexagonal SiPM. This design provides two depth-of-interaction (DOI) layers as well as a high-granularity for event’s positioning. We first present a DOI separation method based on a dispersion metric computed
from the per-channel light distribution, enabling crystal pixel specific calibration for each layer independently. After precise
vertex identification, energy calibration is then performed per
crystal pixel by fitting four reference points: the 202 and 307
keV intrinsic 176Lu peaks and the 511 and 1275 keV peaks
from a 22Na source to a non-linear model accounting for SiPM
saturation. These points are fitted to a non-linear response
model that accounts for SiPM saturation and electronic nonlinearity.
The proposed calibration produces stable photopeak
positions across the detector and achieves an energy resolution
of 15.2% at 511 keV, while preserving the DOI-resolved flood
maps required for high-granularity event’s positioning by PET
imaging techniques.

Abstract

Inter-crystal scatter (ICS) events in PET, in which a 511 keV photon deposits energy across multiple crystals, offer a route to higher effective sensitivity but are conventionally discarded by photopeak windowing. Prior work has focused on positioning the first-interaction crystal, yet positioning alone does not translate ICS recovery into a system-level sensitivity gain: accepting ICS events requires lowering the lower-level discriminator, which sharply increases the singles rate and raises three issues that have not been jointly addressed, how to group raw singles into candidate ICS events, how to reject uncorrelated groupings of independent-photon singles, and whether the resulting triple coincidences are compatible with the dual-coincidence random-correction framework. This study presents a unified pipeline that addresses all three. Candidate ICS events are formed from list-mode singles and screened for type-1 randoms (false ICS) by a joint timing–distance gate, suppressing the singles-level ICS-to-PE ratio from 15.5–18% to 7–7.5% with negligible activity dependence. Validated PE and ICS events are then paired into PE–PE and PE–ICS prompt coincidences, and the singles-prompt (SP) framework is extended to ICS events by symmetric deflation of the measured ICS rate, yielding a type-resolved SP estimator for type-2 random correction. Validated through GATE simulations of the Pharos PET scanner (Brightonix Imaging; Seoul, Korea), the type-resolved SP estimator achieves R̂/R_true = 0.97–1.01 for PE–ICS coincidences across six activity conditions (200–667 μCi), outperforming the unified strategy and standard delayed-window and singles-rate methods. The complete pipeline delivers approximately 13% additional net true coincidences over a PE-only baseline, with essentially activity-invariant gain across the operating range examined.

Abstract

Conventional PET coincidence sorters focus on emission events, disregarding intrinsic Lu-176 background radiation in LYSO (Lutetium-Yttrium Oxyorthosilicate) crystals as unwanted noise. To enable simultaneous acquisition of emission and intrinsic transmission data, we developed a software-based multi-window coincidence processor capable of extracting mixed-energy coincidence events from list-mode singles data. The processor utilizes geometric constraints and a dynamic time-of-flight (TOF)-based limit to filter candidate pairs. To suppress severe 511 keV’s backscatter contamination in the 202 keV transmission window, a dual-condition rejection strategy was integrated, reducing residual crosstalk to 15.9%. Quantitative validation against Monte Carlo simulation ground truth showed good agreement in count statistics, sinograms, and radial profiles. Tomographic reconstructions successfully recovered phantom structures, validating the fidelity of the simultaneously extracted emission and intrinsic transmission coincidences.

Abstract

Positron Emission Particle Tracking (PEPT) enables tracking of individual radiolabelled particles and is widely used in industrial applications (e.g. fluid dynamics), but remains underexplored for biomedical use. We recently demonstrated whole-body in vivo tracking of single nanoparticles in mice using a preclinical PET scanner, showing transit through the vasculature from the injection site. However, murine blood flow varies from a few to several hundred mm/s. Within this range, the conditions under which PEPT can reliably resolve motion remain poorly characterised. Trajectory estimation depends on a complex interplay between activity, particle velocity, acquisition settings (e.g. energy window, temporal sampling), and algorithm parameters, which have not been explored for murine blood flow estimation.

We address this by establishing a controlled framework to evaluate PEPT performance on a 3-ring PET/CT system (Mediso nanoScan 123S). A modified FLSUN 3D printer generated programmable trajectories of Ga-68 radiolabelled beads (0.4–350 kBq) at speeds of 6–170 mms⁻¹. Ground-truth motion was validated via high-frame-rate video and frame-by-frame analysis. Trajectories were reconstructed using the Birmingham method; estimated speeds and positions were compared to ground truth to quantify accuracy and localisation error.

Our results show that tracking a single particle at speeds up to 50 mms⁻¹ with RMSE <2.5 mm requires activity >1.2 kBq. At ≥0.7 kBq, speeds had 18–34% error with large RMSE. At ≤0.5 kBq, trajectories were not reliably resolved. Localisation error increased with decreasing activity and increasing speed, exceeding 10 mm RMSE in low-activity, high-speed regimes. These findings suggest activities >1.2 kBq are required for reliable intravascular speed estimation.