Comprehensive evaluation of a deep learning-based synthetic CT model for MR-only radiotherapy across multiple anatomical sites.

Background and purpose: With advances in deep learning, MRI-only radiotherapy planning (MROP) has been increasingly adopted. However, clinical implementation still requires rigorous evaluation of synthetic CT (sCT) accuracy across multiple anatomical sites.This study aimed to develop and evaluate an all-sites deep learning model for generating sCT from MRI across the head and neck, thorax, abdomen, pelvis, and spine, with comprehensive assessment of image similarity, dose/volume metric fidelity, and patient positioning accuracy. Methods: A modified weakly-supervised learning model was trained to generate sCT, incorporating bone-specific and bladder-specific loss term to improve sCT accuracy. Evaluation included: 1. image similarity analysis between sCT and CT using mean absolute error (MAE) and mean error (ME) of Hounsfield Units (HU); 2. dose/volume metric accuracy assessment by recalculating clinical plans on sCT and comparing dose distribution and dose-volume histogram metrics with those calculated on CT; 3. positioning accuracy assessment, including translational and rotational differences, based on rigid CBCT-to-CT versus CBCT-to-sCT registration. Results: The average MAE in HU was lowest in soft-tissue regions and highest in bone. Across all sites, planning target volume mean dose differences between CT- and sCT-based plans were < 1 Gy. Gamma passing rates exceeded 95% under 2 mm/2% criteria for all regions excluding thorax. Positioning accuracy was highest in head&neck and spine, whereas the abdomen and pelvis regions showed greater variability due to soft-tissue motion and MRI-related artifacts. Conclusions: The all-sites sCT model achieved clinically acceptable image quality, dose/volume metric accuracy, and alignment performance across multiple anatomical regions.