Time-resolved PET imaging using a joint dynamic reconstruction and motion estimation framework (DREME-PET)

Positron emission tomography (PET) plays a critical role in image-guided radiotherapy by enabling accurate tumor localization and contouring. To account for respiration-induced anatomical motion, time-resolved PET imaging is highly desirable, as it captures dynamic anatomical variations to facilitate both disease diagnosis and treatment. However, reconstructing time-resolved PET images is highly challenging due to the severely limited counts available within each short temporal frame (∼0.5 s) and the intrinsic ill-posed nature of the inverse problem.

We proposed a time-resolved PET imaging technique based on a dynamic reconstruction and motion estimation framework (DREME-PET). The framework enables time-resolved volumetric PET imaging and respiratory motion tracking directly from low-count PET list-mode data.

From a conventional PET scan, DREME-PET reconstructs time-resolved PET images while simultaneously generating a patient-specific, data-driven motion model. Based on the solved image and motion model by the time-resolved reconstruction, DREME-PET also enables real-time PET and motion prediction based on subsequently acquired, minimal list-mode data. Specifically, DREME-PET employs spatial implicit neural representation (INR) to represent a reference volumetric PET image, and a corresponding low-rank motion model to map the reference PET image to the time-resolved PET sequence. The motion model comprises a trainable B-spline-based interpolant to represent the low-rank motion basis components (MBC) and a CNN-based motion encoder to derive the MBC scores from time-resolved, low-count list-mode data. In addition to the count and motion variations observed in the training PET scan, the CNN encoder was trained with additional count- and motion-augmentation to render it generalizable to a broader range of motion amplitudes and count variations (to account for the activity decay). DREME-PET are evaluated with an XCAT simulation study, an in-house physical phantom measurement, and a patient study.

DREME-PET allows accurate time-resolved PET reconstruction and motion modeling from a single PET scan, and enables real-time image and motion predictions from subsequent list-mode data acquisitions at a fast inference speed (28 ms per frame). For the XCAT simulation study, DREME-PET achieves an average (±s.d.) image contrast relative error of 14.1% ± 3.0% and a tumor center-of-mass tracking error of 1.7 ± 0.8 mm. The physical phantom study shows an image contrast relative error of 7.4% ± 6.2% and a target center-of-mass tracking error of 0.5 ± 0.5 mm. The patient study shows a Pearson correlation higher than 0.91 and image contrast relative error of 3.6% ± 0.4%.

DREME-PET allows accurate time-resolved PET reconstruction and motion modeling using a conventional, pre-treatment-delivery PET scan. Based on the derived motion model, it also enables real-time, intra-delivery PET imaging and motion tracking with minimal list-mode data to benefit applications such as PET-guided adaptive radiotherapy.