PET scatter correction using machine learning techniques

In positron emission tomography (PET) image reconstruction, loss of contrast and incorrect quantification of activity are produced due to the effect of Compton scattering. Thus, scatter correction becomes essential to improve image quality. In this work, a machine learning approach based on supervised learning has been considered for scatter correction in a simulated multi-ring PET system using a cylindrical phantom. Using positional and energy information from both of the photons detected as input data, we are able to label each coincidence as True or Scattered, which is the archetype of a binary classification problem. Several standard machine learning algorithms are investigated in order to check for best performance in classifying events as True or Scattered coincidences. In order to check for consistency, cross-validation is implemented and different metrics are considered for evaluating performance in our classification task. Among all the tested algorithms, Decision-tree based ensemble algorithms appear to achieve the best classification performances with accuracies superior to 90%. Once the models were trained, they were used to calculate the scatter fraction of new acquisitions in order to compare with the true result given by our Monte-Carlo simulation and to discard scattered labeled coincidences in our samples.