Deep Equilibrium for Hyperparameter Estimation in Dynamic PET Reconstruction - Inria - Institut national de recherche en sciences et technologies du numérique
Communication Dans Un Congrès Année : 2024

Deep Equilibrium for Hyperparameter Estimation in Dynamic PET Reconstruction

Résumé

Positron Emission Tomography (PET) is a functional quantitative medical imaging modality with main applications in oncology and treatment of neurodegenerative diseases. The reconstruction step solves a tomographic ill-posed inverse problem with Poisson noise, with particularly high intensity for dynamic protocols. The injected dose, radioactive tracer distribution, and noise level, are also varying across the exams, which requires to devise a sophisticated regularization scheme adapted to the protocol. In this work, we focus on constructing a parametric model of the dose-dependence with a large number of hyperparameters in regularized dynamic PET reconstruction. Our objective is to emphasize the interpretability of the learned parameters. Classical grid search becomes inefficient to explore such a highdimensional space of hyperparameters. This work investigates a new approach based on the Deep Equilibrium (DEQ) framework. Starting from an iterative algorithm, DEQ performs learning on the fixed point of this algorithm seen as a (deep) sequence of parameterized iterations. The dose-dependence model is then built from a series of DEQ estimates obtained for a constant number of detected events. The proposed method is validated on a synthetic 2D+t database derived from acquired data of a dynamic PET protocol.
Fichier principal
Vignette du fichier
MIC_2024.pdf (437.43 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04707560 , version 1 (24-09-2024)

Licence

Identifiants

  • HAL Id : hal-04707560 , version 1

Citer

Alix Chazottes, Emilie Chouzenoux, Jean-Christophe Pesquet, Florent Sureau. Deep Equilibrium for Hyperparameter Estimation in Dynamic PET Reconstruction. MIC 2024 - IEEE Medical Imaging Conference, Oct 2024, Tampa (Florida), United States. ⟨hal-04707560⟩
0 Consultations
0 Téléchargements

Partager

More