Reconstruct gamma-ray interaction position for the development of an on-chip PET system using deep learning

Ziel/Aim Organoids, stem-cell-derived three-dimensional tissue cultures, find increasing applications ranging from disease modeling to drug discovery and personalized medicine. These growing numbers of uses lead to strong demand for novel measurement capabilities. In this abstract, we present the first steps of developing an on-chip PET system capable of imaging organoids. Here we aimed to prove the concept of improving the reconstruction of the gamma-ray interaction position using deep learning methods.

Methodik/Methods For this purpose, a tentative detection block was designed using a continuous LYSO crystal and silicon photomultiplier (SiPM), whose geometry fits with the microfluidic chips of 3D cell culture. Monte Carlo simulations of this detection block were established on the Geant 4 platform. A large dataset of simulated light pattern images of a wide range of gamma-ray incidence positions and angles were simulated. A CNN based reconstruction network was trained to learn the nonlinear relation between gamma-ray interaction positions and their resulting surface light patterns.

Ergebnisse/Results Various experiments have been run to determine the optimal number of surfaces needed to reconstruct the interaction position from the surface light patterns. Subsequent experiments were then used to find the best CNN backbone architecture for the reconstruction network. The resulting network achieved a mean average error (MAE) of 0.6 mm when trained on a dataset of 100,000 samples and tested on 10,000 samples.

Schlussfolgerungen/Conclusions These preliminary results indicate a promising direction for deep neural network-based methods for gamma-ray interaction position reconstruction in continuous crystals. With a larger dataset and an extensive hyperparameter search, the results will be further improved. In successive experiments, the results achieved with simulated data will be compared to experimental data.

Publication History

Article published online:
08 April 2021