Convolutional Neural Networks in the Classification of Multiphoton Coincidences in a J-PET Scanner.

This work describes an investigation into the utilization of convolutional neural networks for the classification of three-photon coincidences, focusing specifically on the para-positronium decay associated with a photon from nuclear deexcitation. The data were simulated using the Monte Carlo method, with scandium-44 as the source of beta(+) decays. A preprocessing method that allowed for initial cleaning of the training data was described. Subsequently, the parameters of the method for transforming tabular data into images were optimized. Finally, the created images were used to train a binary classifier using a convolutional network model. The developeddata preprocessing step andtransformation method into image format enabled the achievement of a precision rate of 52% at a sensitivity level of 95%, which was a 10 percentage point improvement compared to the logistic regression model.