Towards trustworthy system latency optimization of Positron Emission Tomography with RBC and GNN

Abstract

Positron Emission Tomography (PET) is a highly sensitive imaging instrument, and its system latency plays a pivotal role in real-world applications. We propose a novel trustworthy optimization method to reduce PET system latency by leveraging Role-Based Collaboration (RBC) and Graph Neural Networks (GNN). First, we establish an end-to-end physical model for latency computation, with specific consideration of PET’s characteristics. Then, the GRA (Group Role Assignment)-PET model is proposed based on the E-CARGO (Environments–Classes, Agents, Roles, Groups, and Objects)/GRA model to comprehensively manage the task offloading strategy between PET detector modules and GPU servers, incorporating mechanisms for network contention resolution and computational resource arbitration, with a proof of algorithm convergence provided. Finally, we design an Extended Integer Linear Programming (x-ILP) solution to the proposed model. Furthermore, we develop a trustworthy GNN-based event-filtering approach, leveraging the structural characteristics of PET detector modules. Simulation results show that the proposed collaborative solution significantly improves the overall system performance.