In the field of nuclear medicine, Positron Emission Tomography (PET) is an essential imaging method for diagnosis in oncology, neurology and cardiology. During the examination, positron-emitting radioactive atoms are injected into the patient in the form of a radiopharmaceutical. The decay of the radiotracer releases positrons, each of which annihilates into two gamma photons emitted back-to-back, reaching a pair of opposing detectors surrounding the patient. These coincidentally detected photons are used to reconstruct the distribution of the tracer in the subject's body. The accuracy of this imaging therefore depends on the ability of the detectors to localize and date the interactions of the gamma photons in the scintillators. Recently, the ClearMind project proposed an innovative PET detector concept based on monolithic lead tungstate (PbWO4) crystals sensitive to gamma photon position. This scintillator offers superior sensitivity to traditional pixelated detectors, but introduces additional complexity in signal processing for the spatial reconstruction of interactions.
In this study, the consortium of researchers from the three laboratories of ISAS, DPHP and SHFJ developed an innovative Machine Learning method to improve the localization of gamma photon interactions in the scintillators developed by the ClearMind project. The team used a Density Neural Network, trained on simulations of the detector, with a loss function incorporating the detector's physical constraints (notably near the edges), as well as an estimate of the uncertainty inherent to the reconstruction process. This unique combination has resulted in more robust and reliable position estimates, and the results obtained demonstrate the effectiveness of the proposed approach by highlighting the significant advantages of estimating uncertainties by the model.
By achieving their goal of more reliable and accurate reconstruction of the spatial coordinates of gamma interactions for PET imaging, the authors discuss the fact that their method could be applied to other fields requiring complex sensor data analyses. They predict that their approach could transform the way spatial reconstructions are performed, paving the way for expanded applications as well as more accurate and reliable medical imaging in conjunction with trusted AI.
Contacts and affiliations :
Geoffrey Daniel geoffrey.daniel@cea.fr - Laboratoire d'Intelligence Artificielle et de sciences des Données (SGLS/DM2S/ISAS/DES) ;
Dominique Yvon dominique.yvon@cea.fr - Département de Physique des Particules (IRFU/DRF) et Laboratoire BioMaps (SHFJ/JOLIOT/DRF)