Toward a Unified, Site-Agnostic Deep Learning Reconstruction Framework for Nonstop Gated CBCT In Respiratory Gating Radiotherapy
Abstract
Purpose
Nonstop gated CBCT (ngCBCT) was developed to overcome the limitations of current gated CBCT (gCBCT), enabling 2-8x faster acquisitions and 2.5-3.5x lower imaging dose. However, ngCBCT produces highly non-uniform and under-sampled projection data that challenge conventional reconstruction methods. We developed a deep-learning-based dual-domain-convolutional-neural-network (DDCNN) for ngCBCT reconstruction and evaluated its performance across three respiratory motion-affected disease sites: lung, pancreas, and liver.
Methods
Half-fan gCBCT scans were retrospectively collected from 23 lung cancer patients (54 scans), 13 pancreatic cancer patients (58 scans), and 12 liver cancer patients (52 scans). A total of 328 ngCBCT datasets were emulated by generating two ngCBCT scans from each gCBCT acquisition. Three site-specific DDCNN models were trained using the corresponding datasets. In addition, a unified cross-site DDCNN was trained using ngCBCT scans evenly distributed across the three sites. Model performance was evaluated using both site-specific and cross-site testing, with qualitative visual assessment and quantitative metrics.
Results
Despite substantial differences in projection characteristics and anatomy between thoracic and abdominal regions, DDCNN trained on data from a single disease site (e.g., lung) generalized effectively (PSNR > 33dB and SSIM > 93%) to other sites (e.g., pancreas and liver), producing high-quality reconstructions. Furthermore, a single unified, site-agnostic DDCNN achieved image quality comparable to site-specific models across all evaluated sites, both qualitatively and quantitatively.
Conclusion
The strong cross-site generalizability of DDCNN indicates that the projection-domain network primarily learns acquisition- and physics-driven sampling deficiencies inherent to ngCBCT, rather than site-specific anatomical features, while the image-domain network suppresses residual artifacts and enforces structural consistency. This unified, site-agnostic reconstruction framework represents a key innovation by eliminating the need for disease-site–specific models, thereby simplifying clinical deployment. The integration of ngCBCT with a fast, robust, and generalizable reconstruction strategy has the potential to substantially improve clinical workflow efficiency in respiratory gating radiotherapy.