Real-time multi-resolution image reconstruction for MR-guided prostate radiotherapy

Abstract

Purpose or Objective Intra-fractional tumor motion is a major source of uncertainty in prostate radiotherapy. Using hybrid MR-Linac systems, 3D CINE MR images can be acquired during radiation treatment, enabling real-time motion monitoring and dose accumulation. Typically, a trade-off must be made between frame rate and resolution, rendering it unfeasible to do both simultaneously. We present a novel imaging sequence, dual-density CArtesian SPiRal (dd-CASPR), allowing for image reconstructions at multiple time and spatial resolutions. By coupling this with an in-house developed online reconstruction pipe-line (ReconSocket), we demonstrate the feasibility of three parallel image streams for real-time (i) bulk motion detection (3Hz), (ii) motion quantification (0.8Hz), and (iii) high-resolution images (0.15Hz) of the prostate. Materials and Methods Experiments were performed on a 1.5T Unity MR-linac (Elekta AB, SE) using a 4DMRI motion phantom (ModusQA, CA) and on two healthy volunteers (IRB approved). T1w and bSSFP (T2/T1w) GRE scans were performed using the dd-CASPR sampling pattern (FOV 275x450x110 mm3). k-space profiles were streamed continuously using ReconSocket to a reconstruction server running open-source reconstruction software (BART), generating three different image streams: (i) 1D cranial-caudal (CC) projection navigator at 3Hz, (ii) 3D 10mm3 at 0.5Hz, and (iii) 3D 2mm3 at 0.15Hz. ReconSocket’s streaming latency and jitter was estimated by comparing the 1D navigator stream with the zero-delay phantom output. For the two (ii and iii) 3D streams, the resolved CC motion was estimated using rigid registration on the prostate. Volunteer 1 was instructed to move at the end of the scan to simulate bulk motion. Motion traces from the three image streams were compared to each other. Results The streaming latency of the k-space profiles via ReconSocket to the external reconstruction server was 7ms with a maximum jitter of 10ms, while the reconstruction times of (i-iii) were 25ms, 0.8s, and 7s, respectively. Fig. 1 shows that prostate motion can be tracked on the low-res images (green), but that fast movements are lost on the high-res images (orange). Bulk motion can be clearly detected on the 3Hz projection navigator. Conclusion The proposed 3D dd-CASPR MRI sequence enables real-time reconstructions of a single MRI acquisition at multiple different spatio-temporal resolutions, making it a suitable one-fits-all candidate for online bulk motion detection, target motion monitoring, and dose accumulation.