Deep Learning Enabled Freewill Gated CBCT for Respiratory Gating Radiotherapy
Abstract
Purpose
In respiratory gating radiotherapy (RG-RT), pretreatment imaging—particularly gated cone-beam CT (gCBCT)—is essential but operationally inefficient. Current gCBCT on C-arm linear accelerator is time-consuming (2–8 minutes) and often requires re-scans when gating thresholds need adjustments. To overcome these limitations, we propose the FreeWill gated CBCT paradigm that fundamentally rethinks gCBCT acquisition by decoupling image acquisition from selecting gating thresholds. Specifically, FreeWill gated CBCT performs a 1-minute 3D CBCT scan while simultaneously recording respiratory motion using an external reflector block, enabling retrospective selection of arbitrary gating thresholds and rapid reconstruction of corresponding gated CBCT images.
Methods
FreeWill gated CBCT reconstructs images using only projection data within a selected gating window, discarding projections outside the threshold, resulting in highly non-uniform and under-sampled projection patterns. We developed a dual-domain convolutional neural network (DDCNN) that jointly leverages projection- and image-domain information to reconstruct high-quality images. Feasibility was evaluated using two datasets. First, a half-fan thoracic DDCNN was trained and tested on 77 scans from 30 lung cancer patients treated with respiratory gating. Second, a full-fan thoracic DDCNN was trained using 47 scans from 13 patients and tested on an external, emulated FreeWill gated CBCT dataset (14 scans from 4 patients) to assess cross-institutional generalizability.
Results
For FreeWill gated CBCT with non-uniformly under-sampled projections, the half-fan thoracic DDCNN consistently outperformed conventional reconstruction techniques and image-domain-only CNNs, yielding high-quality images with substantially reduced artifacts. The full-fan thoracic DDCNN similarly demonstrated robust reconstruction performance and strong generalizability across institutions.
Conclusion
We present a novel “scan first, gate later” FreeWill gated CBCT framework that transforms respiratory-gated imaging from a threshold-dependent acquisition into a flexible, efficient, and retrospective process. Combined with rapid DDCNN-based reconstruction (<1 minute on single GPU), FreeWill gated CBCT has the potential to replace current gCBCT, dramatically streamline pretreatment imaging, and improve efficiency in RG-RT.