Holistic evaluation of a machine learning-based timing calibration for PET detectors under varying data sparsity

Area of Science:

• Medical Imaging
• Machine Learning in Physics
• Nuclear Instrumentation

Background:

• Modern PET scanners utilize Time-of-Flight (TOF) information to enhance image Signal-to-Noise Ratio (SNR).
• Accurate timing calibration is crucial for TOF PET, but traditional methods have limitations in correcting higher-order skews.
• Machine learning (ML) shows promise for improving timing resolution but requires extensive calibration times, hindering clinical use.

Purpose of the Study:

• To investigate the impact of data sparsity on ML-based timing calibration for PET scanners.
• To accelerate ML timing calibration for in-system application and clinical feasibility.
• To demonstrate the applicability of ML calibration to analog readout technology.

Main Methods:

• Experimentally acquired calibration data were modified for statistical and spatial sparsity to simulate reduced measurement times.
• Eighty decision tree models were trained with consistent hyperparameters on sparse data and tested on rich data.
• A holistic evaluation framework assessed models using data scientific, physics-based, and PET-based criteria.

Main Results:

• ML-based timing calibration time was reduced from days to minutes without compromising quality.
• Timing resolution significantly improved from (304±5) ps to (216±1) ps compared to analytical methods.
• The ML calibration method proved effective on detectors with analog readout technology.

Conclusions:

• This study presents a significant acceleration of ML-based PET timing calibration, making it suitable for in-system application.
• The ML approach demonstrates versatility by successfully calibrating analog readout detectors.
• The proposed holistic evaluation criteria can guide future ML calibration research in PET imaging.