RSS-Feed abonnieren
DOI: 10.1055/s-0040-1708150
Machine Learning-based Calibration of (Semi-)Monolithic Detectors enabling Depth of Interaction-encoding and Time-of-Flight Capabilities in Clinical PET Systems
Publikationsverlauf
Publikationsdatum:
08. April 2020 (online)
Ziel/Aim Currently available clinical PET systems employ detectors where the scintillator is structured into single needles of 3 – 6 mm width. This limits the spatial resolution and causes radial astigmatism at off-center positions as the depth-of-interaction (DOI) cannot be measured. Introducing more finely structured scintillators with DOI capabilities significantly increases the cost and reduces the sensitivity. We present an alternative based on an array of semi-monolithic crystals which are large blocks without segmentation at costs comparable to current detectors. In this design, optical photons travel in the scintillator and create a 3D-dependent light pattern on multiple photosensor channels enabling the position estimation of the 511 keV gamma photons.
Methodik/Methods An array of 8 LYSO slabs of dimensions 3.9 × 32 x 19 mm3 was coupled to a 64-channel photosensor of digital SiPMs (DPC 3200-22-44, Philips Digital Photon Counting). We employed the supervised machine learning technique Gradient Tree Boosting (GTB), which consists of an ensemble of simple binary decision trees. The algorithm utilizes the raw light pattern of the photosensor. Training data for the machine learning process were obtained using a dedicated fan beam collimator irradiating the detector at known positions. We further determined the coincidence resolving time and energy resolution of the detector after dedicated 3D-position-dependent calibrations.
Ergebnisse/Results We achieved an in-plane spatial resolution of 2 mm FWHM and 3.4 mm FWHM for positioning in DOI with calibration of less than 1 h. An energy resolution of better than 11 % and a coincidence resolving time (CRT) of 280 ps were reached.
Schlussfolgerungen/Conclusions The presented semi-monolithic detector offers an attractive overall performance – including DOI-encoding – by the use of machine-learning algorithms. The concept is a promising alternative to currently employed detector designs improving the spatial resolution of the entire field of view of clinical PET systems.