In 2024, researchers from the Collaborative Computational Project in Synergistic Reconstruction for Biomedical Imaging launched the PET Rapid Image Reconstruction Challenge, or PETRIC, described in Frontiers in Nuclear Medicine as the first challenge ever organised for PET image reconstruction. Challenges have become established instruments for accelerating research in image registration, segmentation and reconstruction in other medical imaging fields, but PET had never had one. The premise was elegantly simple: give competing teams the same reconstruction problem, the same software and the same hardware, and reward whichever algorithm reaches a known solution in the shortest wall-clock time. Because all algorithms used the same projectors and objective functions through the Synergistic Image Reconstruction Framework and the Core Imaging Library, performance differences could be attributed purely to algorithmic approach rather than to software engineering advantages.

The technical core of the challenge was a maximum a-posteriori reconstruction problem, also known as penalised likelihood estimation, using the smoothed relative difference prior. This prior penalises relative differences between neighbouring voxels, preserves edges, and includes voxel-dependent weights derived from the Hessian of the log-likelihood at an initial ordered subsets expectation maximisation reconstruction. Participants had to minimise a well-defined objective combining a Poisson log-likelihood of the measured data, including detection efficiency and attenuation factors, randoms and scatter background, with the regularisation term. The organisers computed gold-standard reference solutions using a modified BSREM algorithm with an EM-like preconditioner, subsets and a relaxed, decreasing step size, a scheme proven to converge to the true MAP optimum. Reaching that converged solution often required on the order of fifteen thousand updates, underscoring just how computationally heavy the reference problem was.

Seven teams registered and four submitted entries, contributing nine algorithms in total. All were gradient-based iterative methods, but they spanned a remarkable range of optimisation theory. Teams deployed stochastic gradient descent with variance reduction techniques such as SAGA and SVRG, which had only recently been adapted to PET, along with the Katyusha acceleration scheme. They experimented extensively with preconditioning, from preconditioners motivated by MLEM to diagonal approximations of the prior Hessian and combinations of the two. Step-size strategies included a-priori decaying schedules, the Barzilai–Borwein rule, Armijo line search and the parameter-free Distance over Weighted Gradients method. Some teams used deep learning to compute warm starts for their iterative schemes, and others fine-tuned subset selection, agreeing that the optimal number of subsets, typically in the range of twenty to fifty, should be scanner specific and depend on factors such as the prime factors of the number of views. Some teams even seeded the randomness of their stochastic algorithms, making them in practice deterministic.

The datasets were as important as the algorithms. The organisers curated phantom acquisitions from a wide range of clinical scanners, including the Siemens Biograph mMR and Vision 600, GE Discovery MI and Discovery 690, the Mediso AnyScan and the compact Positrigo NeuroLF, using phantoms such as the NEMA image quality phantom, the Hoffman brain phantom, an ACR phantom and an Alderson thorax, with radionuclides including fluorine-18, zirconium-89 and germanium-68. Training data were released during the challenge, while evaluation data were acquired at different sites afterwards and released only after the submission deadline. Nobody, including the organisers themselves, had access to the final ground truth data at the time of the competition, which minimised bias toward any vendor or scanner model and ensured fairness. All datasets were made openly available under a Creative Commons license once the challenge formally completed.

Scoring was designed with statistical care. Each algorithm was executed on the standard platform, a 32-core CPU with 100 gigabytes of RAM and an NVIDIA A100 40-gigabyte GPU, until its solution satisfied convergence thresholds on all selected metrics for ten subsequent updates or its runtime exceeded one hour. The metrics were built from the voxel-wise root mean squared error against the converged BSREM reference image, evaluated over the whole object, the background and specific volumes of interest such as tumours, spheres and white and gray matter, together with the absolute error of the mean radioactivity value in each region. Thresholds were chosen to be well below clinically accepted variation while avoiding excessive stringency. Because wall-clock timing and stochastic algorithms introduce variability, each reconstruction was run three times and the median time was used for ranking, with ranks averaged across metrics and datasets.

The results delivered a subtle lesson about algorithm design. While most submitted approaches appear very similar in nature, their specific implementations led to a range of algorithmic performance. Team rankings remained consistent across all runs, yet there was significant variability among individual submissions from the same team, highlighting how difficult it is to maintain performance across heterogeneous datasets and how much can be gained from accurate parameterisation for each scanner. The winning team’s algorithms outperformed the rest overall, and teams that used the same SVRG variance reduction core, such as MaGeZ, UCL-EWS and SOS, achieved different results purely through their choices of preconditioning, subset selection and step sizes. Some algorithms performed well on certain datasets while struggling with others due to poor parameterisation, a finding with direct implications for how reconstruction methods should be tuned in clinical practice.

The outcomes extended well beyond the leaderboard. All three top-scoring teams made their code publicly accessible, providing open-source implementations that others can explore and reuse, and Docker images containing all the necessary software were released to support reproducibility. The curated phantom data collection, the automated computational framework and the backend infrastructure were all released as open-source software, so researchers can now reuse the framework for evaluating new and existing reconstruction methods on new or existing datasets from any PET scanner, radionuclide or phantom. Organising the challenge also exposed and fixed weaknesses in the organisers’ own software, from slow algebraic operations to data array handling in Python, and produced numerous scripts for data ingestion, validation and visualisation. The challenge concluded with a focused workshop at the 2024 IEEE Nuclear Science Symposium and Medical Imaging Conference in Tampa, Florida, where the top-ranked teams presented their methods.

The organisers acknowledge limitations and are already looking ahead. The relatively low turnout may reflect the tight timeline from announcement to conclusion within a single year, as well as partial availability of training data at the start. Converting vendor data into a common open format compatible with STIR and SIRF required closed-source tools available only under research agreements, which restricted the number of contributing sites, though the new emission tomography standardisation initiative and its proposed PETSIRD raw data format may open the door to more scanners in the future. The focus on classical optimisation also made the challenge less suited to deep-learning reconstruction, and the community aims to extend the software with memory-efficient integration of Python CUDA libraries such as PyTorch before organising a deep-learning edition that would evaluate image quality metrics like lesion detectability, contrast and noise rather than convergence. A successor, PETRIC2, using data at reduced count levels, is already open, and adaptations for SPECT reconstruction are being considered.

What makes PETRIC significant is the translational step it narrows. By evaluating algorithms on real measured data rather than simulations, the challenge assessed performance under the noise, heterogeneity and acquisition variability that clinicians actually encounter, making the path from scientific development to clinically relevant application shorter. Reconstruction speed is itself a quality metric of how useful an algorithm is for clinical work, and the challenge demonstrated that careful choices in preconditioning, step sizes and subset selection can yield meaningful gains within an existing framework. As the first challenge of its kind for PET, PETRIC has established foundations that will allow the community to keep benchmarking faster and better reconstruction methods, and with variant versions already launching, the race to make PET imaging quicker is only beginning.

Subject of Research: PET rapid image reconstruction challenge (PETRIC)

Article Title: PET rapid image reconstruction challenge (PETRIC)

Article References: da Costa-Luis, C., Ehrhardt, M. J., Kolbitsch, C., Ovtchinnikov, E., Pasca, E., Thielemans, K., & Tsoumpas, C. (2026). PET rapid image reconstruction challenge (PETRIC). Frontiers in Nuclear Medicine, 6, Article 1770002. https://doi.org/10.3389/fnume.2026.1770002

Image Credits: AI Generated

DOI: 10.3389/fnume.2026.1770002

Keywords: rapid, image, reconstruction, challenge, PETRIC, scientific research