The reduction of ¹⁷⁶Lu background in Lu-based PET scanners using optimized classification

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Machine Learning

Background:

• Lutetium-based (Lu-based) positron emission tomography (PET) scanners are valuable in nuclear medicine.
• Imaging very low activity distributions (<100 kBq) is limited by intrinsic ¹⁷⁶Lu radiation from scintillators, necessitating background reduction.

Purpose of the Study:

• To develop and validate a classification method for distinguishing true coincidences from ¹⁷⁶Lu background radiation in Lu-based PET scanners.
• To enhance the capability of Lu-based PET scanners for imaging very low activity distributions.

Main Methods:

• Investigated five supervised learning classifiers: logistic regression, support vector machine, random forest, extreme gradient boosting (XGBoost), and deep neural network.
• Extracted five energy and time-of-flight (TOF) related features for coincidence event classification.
• Trained the classification model using simulated data and verified its feasibility on measured data from TOF-PET detector modules.

Main Results:

• The extreme gradient boosting (XGBoost) classifier achieved the highest accuracy (>99%) in distinguishing true from background coincidences.
• Imaging tests showed an 89.4% contrast enhancement for a Derenzo phantom and a 52.4% peak-to-valley ratio improvement for a bar phantom at low activity concentrations.
• The method successfully demonstrated improved imaging of low-activity sources.

Conclusions:

• The proposed XGBoost-based classification method effectively reduces ¹⁷⁶Lu background radiation in Lu-based PET scanners.
• This technique significantly extends the application of Lu-based PET scanners to imaging very low activity distributions.
• The method holds potential for various molecular imaging tasks requiring detection of low-level signals.