Single-Ended Readout Depth-of-Interaction Measurements Based on Random Forest Algorithm

The depth-of-interaction (DOI) information can reduce parallax errors and improve the spatial resolution of positron emission tomography (PET) imaging. Current DOI detectors, with either single-ended readout design or dual-ended readout design, usually require substantial changes to the crystal arrays and/or readout electronics. In this work, we proposed a single-ended readout DOI estimation approach with minimal changes in the detector hardware design, using a supervised machine learning algorithm, random forest (RF). The energy value measured from one end of the crystal, combined with sample points of the pulse signal, were used as inputs to the DOI estimation algorithm, which provided a DOI performance close to that of the dual-ended readout design. We compared the proposed single-ended RF method with the classic dual-ended energy ratio method and a dual-ended version of the RF method, using positioning error and resolution as performance metrics for DOI estimation. All three methods achieved average absolute DOI positioning errors of < 1.00 mm and average DOI resolutions of < 2.20 mm full width at half maximum (FWHM). More specifically, we achieved the averaged DOI resolutions of 2.15, 1.80, and 1.93 mm FWHM for the single-ended RF method, dual-ended RF method, and dual-ended energy ratio method, respectively. In conclusion, the single-ended RF method achieved comparable performance without extra photodetectors and readout electronics and may provide a feasible solution for DOI estimation.