Efficient High-Resolution Reconstruction for Limited-Angle Non-Coplanar CBCT Imaging In Non-Coplanar Radiation Therapy
Abstract
Purpose
In non-coplanar SBRT/SRS, efficient high-resolution volumetric imaging under limited-angle non-coplanar CBCT (LA-NC-CBCT) acquisition is critical for avoiding potential collisions while ensuring treatment delivery accuracy. However, severe data occlusion outside the scanning range makes it extremely challenging to produce high-quality volumetric images. Moreover, high-resolution image reconstruction often incurs substantial computational cost, which limits the clinical applicability of advanced imaging algorithms. This work aims to address the challenges of efficient reconstruction of high-resolution images for LA-NC-CBCT scans in a patient-specific manner.
Methods
We propose a prior-knowledge-embedded implicit neural representation (PINR) framework with a dual-domain ordered-subsets (DDOS) based optimization strategy. The PINR network is pretrained using a full-scan CBCT image acquired at an initial coplanar position of a patient. To account for positional discrepancies between coplanar and non-coplanar acquisitions, a motion-encoded regularizer is introduced. It incorporates treatment couch motion information to transform the coplanar prior to the corresponding non-coplanar configuration. Furthermore, we develop a novel DDOS strategy that partitions both image volumes and projections in a specific order based on super-resolution principles. This enables network updates using smaller subsets of data, alleviating GPU memory constraints and reducing the number of epochs, thereby substantially improving computational efficiency.
Results
Experiments on phantom measurement datasets demonstrate that the proposed method consistently outperforms the conventional FDK and the reconstruction-by-registration baselines across a wide range of limited-angle scenarios. The method achieves improved image quality and structural fidelity with stable and monotonic convergence behavior for scanning ranges from 90o down to 15o. In terms of computational cost, the method can generate a high-quality volumetric image of resolution 512x512x94 within one minute on a single 48-GB GPU server.
Conclusion
We propose a novel data- and computationally efficient solution for LA-NC-CBCT image reconstruction, enabling high-quality and robust imaging for non-coplanar SBRT/SRS with the existing onboard CBCT imaging system.