High throughput software-based gradient tree boosting positioning for PET systems.

Area of Science:

• Medical Physics
• Machine Learning
• Computational Imaging

Background:

• Gradient Tree Boosting (GTB) demonstrates high accuracy for gamma interaction position estimation in PET crystals during bench-top experiments.
• Scaling GTB to full preclinical and clinical PET systems necessitates near real-time processing capabilities.

Purpose of the Study:

• To develop and optimize a high-throughput C++ implementation of GTB for singles positioning in PET systems.
• To evaluate the impact of various optimizations on processing throughput.
• To investigate feature and parameter selection for GTB in segmented detectors.

Main Methods:

• A C++ implementation of GTB-based singles positioning was developed.
• Optimizations were systematically evaluated for their effect on processing throughput.
• Feature and parameter selection for GTB was investigated for the Hyperion II^DPET insert using two main models and various hyperparameters.

Main Results:

• The proposed framework achieved singles positioning throughputs exceeding 9.5 GB/s for simpler models and 240 MB/s for complex models on a Skylake server.
• Detailed throughput analysis identified key performance-limiting factors.
• An empirical throughput model was derived with a Mean Absolute Error (MAE) of 175.78 MB/s for throughput estimations.

Conclusions:

• The optimized GTB implementation significantly enhances processing throughput for PET singles positioning.
• The derived empirical throughput model provides valuable guidance for GTB model selection and PET scanner design.
• This work facilitates the transition of GTB to real-time preclinical and clinical PET applications.