MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Convolutional Neural Network for Centroiding and Depth-Of-Interaction Localization in PET

A LaBella*, W Zhao , AH Goldan , Stony Brook University, Stony Brook, NY

Presentations

(Wednesday, 7/17/2019) 7:30 AM - 9:30 AM

Room: 225BCD

Purpose: To explore the feasibility of using convolutional neural networks (CNNs) for enhanced localization of gamma ray absorption and depth-of-interaction (DOI) identification in PET systems.

Methods: PET data was generated using the Monte Carlo software TracePro. Two detector systems were simulated, consisting of 24x24 (1.0x1.0x20mm³ crystals) and 16x16 LYSO scintillator arrays (1.5x1.5x20mm³ crystals), coupled 9-to-1 and 4-to-1 to SiPM arrays with 3x3mm² readout pixels. Light guides were optically coupled to scintillators to enable single-sided DOI readout and enhance centroiding. 25,000 unique 511 keV gamma ray absorption events were simulated. Absorption positions in 3D space were simulated uniformly. CNNs were generated in Keras 2.2.4 and Tensorflow 1.12. 80% of the simulated data was used for model training while 20% was used for testing. 10% of the training dataset was held out for training validation. Hidden layers were followed by ReLU activation. The output layer used linear activation to output 3D coordinates representing gamma ray absorption position. Mean-squared error (MSE) was used to characterize CNN performance. We evaluated the CNN using 20,000 energy-resolved events from experimental flood histogram data obtained from a 4-to-1 coupled detector module using a 3MBq Na-22 source.

Results: MSE in the x-y plane was equal to 0.297mm² for 4-to-1 coupling and 0.1403mm² for 9-to-1 coupling. MSE in the z-direction was equal to 2.13mm² for 4-to-1 coupling, which is consistent with DOI localization performance reported in the literature for similar systems, and 1.33mm² for 9-to-1 coupling. Our CNN was able to correctly centroid experimental flood histogram data. DOI localization followed Beer-Lambert Law for absorption depth.

Conclusion: We were able to demonstrate the potential of CNNs for PET centroiding and DOI identification. We found the CNN model trained with simulated gamma ray interactions transfers well to real PET data, which may potentially enhance centroiding and DOI localization in practice.

Keywords

PET, Spatial Resolution, Scintillators

Taxonomy

IM- PET : Machine learning, computer vision

Contact Email