PET detectors based on monolithic crystals show better performance with respect to the ones based on pixellated crystals in terms of 2-D spatial resolution and depth-of-interaction estimation capabilities.
However, they need a long and complex calibration procedure to reach optimal performance and the best event positioning algorithms are too complex for a real-time implementation in a clinical PET scanner. Artificial intelligence (AI) and in particular neural networks seem to be a solution to this problem, achieving the best performance with relatively low computational complexity. However, these algorithms do not have solid theoretical foundations and are prone to severe over-fitting.
We present an explainability framework for AI algorithms to use in the optimization of PET detectors based on monolithic crystals. The framework is able to show the internal working of the network, allowing to build trust in it and to improve its performance.