Multi-Stage Volumetric Synthetic CT Reconstruction from Diagonal X-Ray Images for Carbon Ion Radiotherapy
Abstract
Purpose
In carbon ion radiotherapy (CIRT), treatment-day position verification is typically performed using two-dimensional (2D) X-ray images acquired at non-orthogonal angles. While effective for bone-based alignment, this approach provides limited three-dimensional (3D) anatomical information, restricting comprehensive assessment of anatomical consistency relative to the planning CT. This study investigates the feasibility of reconstructing volumetric synthetic CT (sCT) images from diagonally acquired verification X-rays for enhanced anatomical verification in CIRT.
Methods
A retrospective dataset of 401 prostate cancer patients treated with CIRT was analyzed. Among them, 109 patients had geometrically aligned planning CT and in-room X-ray data, while the remaining cases contained CT-only data. Verification X-rays were converted to CT-consistent projections and combined with CT-derived reference projections to compensate for the limited field-of-view, enabling initial volumetric reconstruction. A three-stage deep learning framework was applied to progressively reconstruct volumetric sCT images from limited-view X-ray inputs. The method was evaluated on an independent test set of 20 patients and compared with existing approaches using mean absolute error (MAE), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS).
Results
The proposed framework demonstrated superior reconstruction performance compared with existing methods across all evaluation metrics. Improved preservation of bony anatomy and soft-tissue structures was observed, with reduced reconstruction artifacts and more coherent volumetric appearance across axial, coronal, and sagittal planes. Ablation studies confirmed that both slice-wise refinement and volumetric adjustment stages were critical for stabilizing anatomical structure and global intensity consistency.
Conclusion
This study demonstrates the feasibility of reconstructing volumetric synthetic CT images from two diagonally acquired verification X-rays in CIRT. By addressing the inherent limitations of conventional 2D position verification, the proposed framework enables enhanced 3D anatomical assessment using routinely available imaging, without requiring additional hardware or full volumetric imaging systems.