Author: Yu Gao, Lei Xing, Siqi Ye 👨🔬
Affiliation: Department of Radiation Oncology, Stanford University 🌍
Purpose:
Limited-angle CBCT (LA-CBCT) scans are often the only option for non-coplanar radiation therapy to prevent potential mechanical collisions. However, the consecutive angular occlusion of projections in LA-CBCT makes reconstructing high-quality volumetric images extremely challenging, hindering efficient and precise treatment adaptation and delivery. Deep learning methods have shown great promise in sparse-data image restoration but often face challenges due to limited training data. Additionally, the large-scale domain transformation involved in CBCT imaging further hinders their application in clinical practice. This work addresses the challenges of data scarcity and the memory bottleneck by developing a dual-domain ordered-subset (DDOS)-based neural representation network using prior image embedding (NeRP), for accelerated and high-resolution LA-CBCT image reconstruction.
Methods:
We leverage the NeRP framework for LA-CBCT image reconstruction using an 8-layer MLP, trained to overfit a prior reference CBCT volume for initialization. To overcome memory limits, we propose an ordered-subset (OS) partitioning strategy with a “downsample-and-shift” approach for both image voxels and projections. The subset-dependent image geometry and corresponding domain transform were computed dynamically, enabling batch-based optimization and efficient high-resolution reconstruction.
We conducted experiments on a head-and-neck phantom using Varian’s TrueBeam machine. Couch kicks were applied to acquire LA-CBCT projections. A single 24GB GPU was capable to reconstruct high-resolution images of size 94x512x512. Reconstructions with various limited-angle coverages from 90o down to 15o were evaluated against standard analytical reconstruction method FDK using PSNR, SSIM metrics, and visual quality assessment.
Results:
The proposed DDOS-NeRP significantly outperformed the FDK approach in image quality, with improved robustness across various angle coverages. Additionally, DDOS-NeRP achieved convergent results within 20 minutes, and can be further accelerated by increasing the image-domain OS size with a larger memory GPU.
Conclusion:
The proposed DDOS-NeRP framework enables dataset-free and memory-efficient high-resolution reconstruction with improved accuracy, speed, and robustness for LA-CBCT imaging in non-coplanar radiation therapy.