Experimental Characterization of Embeddable Machine Learning Reconstruction Algorithms for Anger Cameras

This work presents the experimental characterization of position sensitivity accuracy when implementing machine learning algorithms for the reconstruction of γ rays interaction points in Anger cameras for medical imaging. The work aims to show state-of-the-art position sensitivity when using an embeddable and fast Decision Tree classifier, combined with Principal Component Analysis. The training dataset acquisition procedure is discussed, together with clustering techniques implemented for the automation of data labeling. To overcome the impossibility to distinguish adjacent classes in the training set with sub-mm distance between classes when using a regular parallel-hole collimator, we implemented two innovative solutions for the training procedure: a custom mechanical structure scanning the module used in experimental characterization, and the use of data augmentation to generate new fractions of training dataset from perturbation of the experimental dataset. Preliminary results have been obtained using the INSERT gamma camera, where a 50×100×8 mm^3 CsI(Tl) scintillation crystal is coupled to 8 tiles of RGB-HD SiPMs, and the charge signal is collected by two 36-channel ANGUS ASICs. The energy resolution equals 17% at 140 keV, and the module is capable to acquire events up to ~10 kcps. Count rate is only limited by typical detector performance (e.g. front-end count rate capability, data transmission bandwidth); in fact, implementation of Decision Tree classification in FPGA (free of data pre-processing) would allow edge-computing of the position sensitivity, with sub-µs data latency. Experimental results show 1.2 mm FWHM spatial accuracy, which is equal to the accuracy obtained with the specific module under test when implementing statistical position reconstruction algorithms such as maximum likelihood estimation.