Unsupervised Depth-of-Interaction Estimation in Monolithic Crystal PET Detectors With Wasserstein Distance-Based...

Monolithic crystal PET detectors are equipped with depth of interaction (DOI) encoding, and neural network (NN) is reported as a useful tool for estimating DOI from light distribution (LD). However, obtaining an LD-DOI paired dataset is challenging when only perpendicular irradiation is available. To estimate DOI only from LD, we introduce a data-driven method called Wasserstein distance-based Locally Linear Embedding, or simply W-LLE, which is applicable to row-column sum signal readout. The effectiveness of W-LLE has been evaluated using both simulated and experimental data. In the simulation study, W-LLE outperformed a model-based method and showed comparable DOI accuracy to the original Locally Linear Embedding (LLE) applied to individual readout setups. In an experimental configuration, DOI resolution was measured indirectly using NNs trained with DOI estimates derived from these methods. Assuming that $x$ -direction and DOI errors are uncorrelated, NNs trained with W-LLE outperformed those trained with the model-based approach, with DOI resolution of up to 1.8 mm FWHM.