Synthetic CT generation from cone-beam CT using deep learning for head-and-neck adaptive radiotherapy

Background and purpose: Synthetic computed tomography generation from cone-beam computed tomography enables adaptive radiotherapy on conventional linear accelerators, but its clinical implementation is limited. We developed and validated a deep learning model for synthetic computed tomography generation in head-and-neck patients, including regions beyond the cone-beam computed tomography field of view, and evaluated its accuracy for adaptive radiotherapy workflows. Materials and methods: A UNet3+ generative adversarial network with multi-component losses was trained on 148 patients (90 training, 38 validation, 20 independent testing). Anatomical concordance, intensity fidelity versus the deformably registered planning computed tomography, and dose recalculation in a commercial treatment planning system (global gamma at 3%/2 mm and 2%/2 mm, 10% dose threshold) were assessed. Results: In Hounsfield units, mean absolute error was 49.9 ± 11.0 globally, 40.2 ± 8.2 inside the cone-beam field of view (mean error 0.2 ± 2.7) and 61.4 ± 19.0 outside. Mean differences in planning target volume dose metrics were < 1.6% (maximum 5.1%), and mean gamma pass rates were 99.0% (3%, 2 mm) and 98.7% (2%, 2 mm). Surface Dice at a 2-mm tolerance was 0.81 ± 0.06 (Dice 0.97 ± 0.01; mean surface distance 1.07 ± 0.31 mm; volume similarity 0.99 ± 0.01). Conclusions: The proposed model generated accurate, anatomically coherent synthetic computed tomography from cone-beam computed tomography, including realistic reconstruction beyond the cone-beam field of view, supporting the feasibility for daily adaptive-radiotherapy dose recalculation on conventional linear accelerators and motivating prospective clinical evaluation.