Improvement of phoswich detector-based β+/γ-ray discrimination algorithm with deep learning

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Machine Learning in Healthcare

Background:

• Positron probes aid malignant tumor localization using radiopharmaceuticals.
• Conventional methods struggle to distinguish gamma rays from positrons, increasing detection errors.
• Existing techniques like multilayer scintillator detection have limitations in accuracy.

Purpose of the Study:

• To enhance positron detection accuracy by analyzing energy distribution in multilayer scintillator detectors.
• To develop an improved algorithm for discriminating true positrons from false ones.
• To reduce misidentification of gamma rays as positrons in tumor imaging.

Main Methods:

• Utilized Autoencoder, an unsupervised deep learning model, for signal processing.
• Trained Autoencoder to separate signals from each scintillator layer.
• Applied energy windowing to energy distribution data for positron discrimination.

Main Results:

• The Autoencoder-generated energy distribution map closely matched simulation results.
• Positron detection sensitivity increased by 29.79% compared to conventional methods at an equal error rate.
• The proposed method achieved a 25.0% lower error rate at the same sensitivity level.

Conclusions:

• Developed an Autoencoder-based algorithm for superior true positron discrimination.
• Demonstrated increased positron detection sensitivity with a maintained low error rate.
• Potential for improved accuracy and speed in cancer localization using advanced probes and cameras.