Non-invasive full quantification of dynamic [11C]PBR28 PET imaging with ASTRA

Abstract Quantitative analysis of dynamic PET imaging requires metabolite-corrected arterial input functions, traditionally obtained through invasive arterial sampling. This limits broader adoption of fully quantitative kinetic modelling, particularly for tracers such as [11C]PBR28 that generate radiometabolites. Here we demonstrate that metabolite-corrected input functions can be estimated from dynamic PET images alone across heterogeneous scanners and acquisition protocols, enabling full quantification without blood sampling. We developed a deep learning framework, ASTRA, that estimates subject-specific arterial input functions from brain PET data while quantifying predictive uncertainty. Across two independent multi-centre datasets acquired with different scanners and protocols, ASTRA-derived estimates closely matched arterial blood–based quantification. In a fully held-out clinical cohort without ground-truth blood data, ASTRA reproduced previously reported semi-quantitative SUVR group differences. Finally, in a fourth dataset, ASTRA preserved sensitivity to pharmacological TSPO blockade. This work addresses a longstanding barrier in dynamic PET imaging and establishes a scalable path toward blood-free, fully quantitative [11C]PBR28 PET imaging.