Self-Organizing Map Neural Network-Based Depth-of-Interaction Determination for Continuous Crystal PET Detectors

For continuous crystal-based PET detectors, not only the two-dimensional (2D) plane coordinate of the interaction point, but also the depth-of-interaction (DOI) of the event could be precisely estimated by the single-end readout of the scintillation light. In this paper, we propose a practical method for DOI determination for continuous crystal PET detectors. By self-organizing map (SOM) neural network with unsupervised learning, the perpendicularly irradiated reference events in each reference position are classified into a certain number of groups, which are simultaneously used for the plane coordinate estimation and the DOI decoding. The reference events measured in an oblique irradiation are used to generate the initial weights of the SOM neurons and to calibrate the DOI decoding. Applying the new method to our experimental data, the SOM-based DOI estimation could achieve an average resolution of 2.56 mm over the whole thickness (0-10 mm) of the crystal. The DOI effect is also evaluated by comparing the plane position resolutions of the test beams in different incident angles with and without the DOI correction. The test results show that the DOI determination could seamlessly integrate into our previously proposed the SOM-based plane coordinate estimation scheme to realize the high performance real-time three-dimensional position estimation for continuous crystal-based PET detectors.