Machine learning positioning algorithms for long semi-monolithic scintillator PET detectors

Area of Science:

• Nuclear Instrumentation
• Medical Physics
• Machine Learning Applications

Background:

• Semi-monolithic scintillator detectors are vital for high-resolution imaging.
• Existing analytical positioning methods have limitations in spatial resolution, particularly at detector ends.
• Improving spatial resolution is key to advancing detector performance in various scientific fields.

Purpose of the Study:

• To develop and evaluate machine learning positioning algorithms for semi-monolithic scintillator detectors.
• To enhance spatial resolution in both the monolithic (y) and depth of interaction (z) directions.
• To compare the performance of machine learning methods against traditional analytical techniques.

Main Methods:

• Manufactured two semi-monolithic scintillator detectors using lutetium yttrium oxyorthosilicate (LYSO) slabs.
• Utilized a 4x16 silicon photomultiplier array for readout.
• Employed Extreme Gradient Boosting (XGBoost) machine learning models, optimized with Genetic Algorithm (GA) or Particle Swarm Optimization (PSO), to predict interaction positions.

Main Results:

• Machine learning positioning methods significantly improved both y and z spatial resolutions compared to analytical methods.
• Achieved average y spatial resolutions of 0.92 ± 0.41 mm and 0.94 ± 0.44 mm, outperforming the 1.38-1.39 mm of analytical methods.
• Attained average z spatial resolutions of 1.67 ± 0.41 mm and 1.68 ± 0.45 mm, superior to the 2.09-2.14 mm from analytical methods.

Conclusions:

• Machine learning algorithms, specifically XGBoost, offer superior spatial resolution for semi-monolithic scintillator detectors.
• These algorithms achieve sub-millimeter y-spatial resolution (<1 mm) and sub-2 mm z-spatial resolution (<2 mm).
• The developed methods represent a significant advancement for high-precision scintillator detector applications.