You are here : Home > News > Reconstructing PET images more reliably with a convergent “Plug-and-Play” regularization approach

Scientific result | Positron Emission Tomography | Artificial intelligence

Reconstructing PET images more reliably with a convergent “Plug-and-Play” regularization approach


A team of researchers from BioMaps (SHFJ) is proposing a new reconstruction method for PET imaging that combines optimization algorithms with deep neural networks. Their “plug-and-play” regularization approach ensures improved image quality thanks to PET-specific reconstruction learning, while guaranteeing stability and robustness—two critical aspects in the use of AI-based medical techniques.​

Published on 16 December 2025

Positron emission tomography (PET) is commonly used in clinical settings to diagnose cancers, neurodegenerative diseases, and cardiac disorders. When the collected data are limited—either due to reduced injected dose, dynamic imaging, or faster acquisition—the reconstructed images become more uncertain, making interpretation more difficult. Classical model-based iterative reconstruction methods are robust but perform poorly when modeling the underlying medical images, resulting in degraded reconstructions under high noise conditions. Conversely, deep learning–based methods can improve reconstructed image quality, but often at the cost of instabilities linked to measurement noise and to the limited size of training datasets.

Une approche hybride

A team from BioMaps (SHFJ) introduces a hybrid approach known as Plug-and-Play (PnP), which leverages the strengths of both classical modeling and machine learning, while scaling efficiently to high-dimensional problems such as PET. An optimization algorithm simultaneously enforces consistency between the reconstructed image and the measured data, as well as between the image and the learned properties of medical images from a training set, all within a mathematically controlled framework. The neural network is trained to guarantee convergence of the entire process, using during training the fixed-point equation satisfied at convergence.

Validated on realistic low-dose simulations and applied to patient data, this reconstruction method is competitive with “classical” model-based approaches and state-of-the-art learning-based reconstruction techniques. It better preserves fine structures, reduces noise in homogeneous regions, and generalizes remarkably well to unseen cases, including simulated lesions or pathological examples absent from the training set for real patient data.

This method offers a promising avenue for reducing PET injected dose without sacrificing diagnostic accuracy. It also opens perspectives for other imaging modalities where stability and image quality are essential.

Contact at Frédéric-Joliot Institute for Life Sciences:

Mari​​on Savanier (marion.savanier@cea.fr)

Floren​​t Sureau (florent.sureau@cea.fr)



Top page