Speaker
Description
Positron Emission Tomography (PET) is a biomedical imaging technique based on the detection of the two collinear 511-keV photons originated from the annihilation of the positrons emitted by administered radiotracers. Before the annihilation occurs, positrons travel some distance, known as positron range (PR). PR depends both on the initial energy of the positrons, which follows a different distribution for each radionuclide, and for the composition and density of the surrounding materials. PR produces a blurring in PET images which is especially relevant in preclinical studies with small animals, where the PR may be of the same size as some relevant structures. In the last few years, a wide number of radiotracers labeled with radionuclides with large PR such as 68Ga and 124I have been proposed. In these cases, PR may significantly compromise the spatial resolution.
There are many different approaches to model and correct the PR in PET. In this work, we propose the use of a Neural Network (NN) to perform the PR correction as a post-processing step applied to the reconstructed images. The NN has been trained with simulated cases of 18F and 68Ga acquisitions from a set of numerical mice models. In order to generate realistic cases, the PR was simulated with penEasy, a Monte Carlo simulation tool based on PENELOPE, and then the PET acquisition was simulated with MCGPU-PET and reconstructed with an iterative MLEM algorithm. A U-Net NN, commonly used in medical imaging, was used to train a model able to convert PR-blurred images to their PR-free counterparts. Our results show that the proposed approach is able to perform PR corrections improving the accuracy of the PET images up to 95% of the original images in a fast and accurate way without incrementing statistical noise.
