Time-Resolved PET Imaging Using a Joint Dynamic Reconstruction and Motion Estimation Framework (DREME-PET)
Abstract
Purpose
Positron emission tomography (PET) is essential for image-guided radiotherapy by enabling accurate tumor localization and delineation. For sites affected by respiration, time-resolved PET is needed to resolve motion but challenged by very low counts per timeframe (~0.5s) and an ill-posed reconstruction problem. We developed DREME-PET, a joint dynamic reconstruction and motion-estimation framework that enables time-resolved volumetric PET imaging and real-time respiratory motion tracking directly from low-count list-mode data.
Methods
DREME-PET simultaneously reconstructs a time-resolved PET sequence and learns a patient-specific low-rank motion model in an ‘one-shot’ framework. From a conventional PET scan, it estimates a reference activity map via implicit-neural-representation and models motion using B-spline-bases with a CNN-encoder. The encoder ingests low-count list-mode data from each timeframe to generate coefficients that compose deformation-vector-fields (DVFs) from the B-spline-bases, which deform the reference map into frame-specific PET images. Once trained, the CNN-encoder can resolve intra-scan motion of the conventional PET scan and also leverage minimal future list-mode signals for real-time motion and image estimation, with robustness enhanced through count- and motion-augmentation. DREME-PET was evaluated using XCAT phantom simulations and physical phantom measurements.
Results
For XCAT, DREME-PET achieves an average(±s.d.) image contrast-relative-error (CRE) of 15.2±1.4% and a tumor center-of-mass-error (COME) of 1.6±0.8mm for dynamic reconstruction, and 14.0±3.3% and 1.8±0.9mm for real-time imaging. By comparison, 4D-PET results are 50.7±7.6% and 6.9±0.3mm. Physical phantom measurements show DREME-PET achieves an image CRE of 7.9±5.7% and a COME of 0.6±0.4mm for dynamic reconstruction, and 7.2±6.3% and 0.5±0.5mm for real-time imaging, versus 8.3±5.1% and 2.7±1.4mm for 4D-PET. For real-time imaging, DREME-PET’s inference time is 28ms for each minimal list-mode signal(<0.5s), meeting real-time temporal constraints.
Conclusion
DREME-PET allows accurate time-resolved PET reconstruction using a conventional PET scan. It also enables real-time, intra-delivery PET imaging and motion-tracking with minimal list-mode data to benefit applications such as PET-guided motion-adaptive radiotherapy.