Perception-Enhanced Generative Adversarial Network for Synthesizing Tau Positron Emission Tomography images from Structural Magnetic Resonance Images: a cross-center and cross-tracer study

Abstract

Abstract Different tau positron emission tomography (PET) radiotracer holds promises for monitoring the accumulation of tau pathology in vivo. However, the low availability relative to the massive demand for tau-PET has ultimately hinders the potential benefits for the majority of patients. Here, we developed a unified deep learning model for synthesizing tau positron emission tomography (PET) images from the more available structural magnetic resonance imaging (sMRI). A total of 1387 subjects from two different cohorts were included in this study, involving tau-PET with 18F-flortaucipir and 18F-florzolotau. The tracer-specific models were trained independently and cross-validated internally and externally. The overall evaluations and regional-specific quantitative evaluations of the synthetic tau-PET have verified that the synthetic tau-PET followed a uniform distribution of reality and could accurately quantifying regional tau deposition, and the proposed method achieved the state-of-the-art performances on commonly used metrics and satisfies the reconstruction needs for clinical standards.