Ponente
Descripción
Compton camera imaging is currently under active investigation in the field of medical physics, with researchers exploring its potential for various applications. The IRIS group at IFIC-Valencia has assembled the MACACO III Compton camera prototype, equipped with three detector planes housing LaBr3 crystals coupled to silicon photomultiplier arrays. This technology aims to accurately determine the distribution of incoming photons. Thus, it necessitates image reconstruction algorithms that rely on the precise positions of photon interactions within the crystals. Previous studies have explored machine learning algorithms and analytical methods to calculate those interaction coordinates. In this work, we focused on the application of a convolutional neural network model, trained with simulated data. The main aim was to predict photon interaction positions within a monolithic scintillator crystal. The training dataset was generated through simulations conducted in Gate v8.0, utilizing a LaBr3 crystal coupled to a silicon photomultiplier array. These simulations captured the light collected by the photomultiplier array pixels for both model training and testing. The convolutional neural network model architecture was adapted from the well-known model VGG16. FWHM, Euclidean distance and Mean Absolute Error were used to assess the model's accuracy. After achieving an effective model for predicting x and y coordinates, its application was extended to estimate the depth of interaction. A comparison with earlier studies highlights the reliable performance of the model in both 2D and 3D research. The CNN model was compared to a state-of-the-art analytical method developed in the IRIS group. The CNN model predictions were found to be promising for accurately localizing photon interactions within a scintillator crystal.
