Single prior-based 3D volumetric tracking from a planar snapshot in spine SBRT: A proof-of-concept study.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiation Oncology, Washington University in St Louis School of Medicine, St Louis, MO, USA; Department of Nuclear Engineering, Hanyang University, Seoul, Republic of Korea; Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, USA.
- Department of Radiation Oncology, Washington University in St Louis School of Medicine, St Louis, MO, USA.
- Department of Nuclear Engineering, Hanyang University, Seoul, Republic of Korea. Electronic address: [email protected].
Abstract
Continuous 3D onboard monitoring is essential to ensure accurate dose delivery of stereotactic body radiation therapy. However, current approaches for 2D-to-3D volumetric estimation typically require extensive patient-specific data, or depend on auxiliary imaging hardware. This study introduces a proof-of-concept framework for data-efficient, patient-specific 3D volumetric tracking using only a single volumetric prior. A retrospective dataset of 80 spine SBRT patients was utilized. Digitally Reconstructed Radiographs synthesized from cone-beam computed tomography volumes served as surrogates for onboard kV projections. The pipeline consists of patient-agnostic pre-training using an archival dataset, patient-specific adaptation using the single volumetric prior supplemented by non-rigid motion simulations, and simulated intra-fractional 3D estimation. The network employs a modified Attention U-Net with a channel-to-depth bottleneck and a triple feature concatenation to map the 2D to 3D spatial domain. The proposed model demonstrated a statistically significant improvement in structural tracking capabilities compared to a static uncompensated baseline. Furthermore, the performance variance across patients was substantially compressed, enhancing reliability against setup uncertainties while achieving high spatial concordance in the complex spinal anatomy. The inference throughput reached approximately 10 frames per second. This proof-of-concept study establishes the technical feasibility of data-efficient, patient-specific onboard volumetric tracking using a single volumetric prior by successfully resolving anatomical deformations from a single 2D projection in near real-time.