Experimental Assessment of PCA and DT Classification for Streamlined Position Reconstruction in Anger Cameras

In this article, we present a study of machine learning (ML) algorithms to simplify the computation of the planar scintillation coordinates in Anger Cameras for emission tomography applications. Two ML-based techniques for data inference and one technique to speed-up the training procedure are explored within the framework of a multimodal SPECT scanner. First, the use of principal component analysis (PCA), a dimensionality reduction algorithm, is explored to reduce the computational complexity of the maximum-likelihood statistical estimation method. The analysis indicates a
$\sim 3$
-fold reduction of computational complexity for typical Anger Camera architectures (with 72 channels). Second, the estimation of the scintillation coordinates is formulated as a classification problem, addressed by means of a decision tree (DT) classifier. No degradation of the achievable intrinsic spatial resolution
$(\sim 1.2$
-mm FWHM) of the detection module was observed when applying PCA (reducing from 72 to 25 components). The DT classifier was trained on experimental data obtained using a parallel-hole collimator: again no degradation of spatial resolution is observed and the computation cost is reduced by more than two orders of magnitude. Finally, in order to overcome the limits of a cumbersome training procedure involving the translation of the collimator, data augmentation was successfully leveraged for the generation of artificial data.