Improved synthetic CT image quality does not guarantee dosimetric agreement in CBCT-based adaptive proton radiotherapy

Abstract

Objective. Synthetic CT (sCT) generation from cone-beam CT (CBCT) has emerged as a promising enabler for adaptive radiotherapy. While deep learning (DL) methods have demonstrated substantial improvements in image quality, their impact on proton beam therapy (PBT) dose calculation remains unclear, particularly for CBCT acquired on proton gantries, which are prone to increased artifacts and noise. Approach. In this study, we evaluated sCT generation from proton gantry CBCT using a denoising diffusion probabilistic model (DDPM) and a cycle-consistent generative adversarial network, and compared them against non-learning-based correction methods (corrected CBCT (corrCBCT) and virtual CT (vCT)) implemented in a commercial treatment planning system. Image quality (mean absolute error, peak signal-to-noise ratio, normalized cross-correlation) and dosimetric accuracy (gamma passing rate/GPR, dose-volume histogram metrics) were assessed in head-and-neck (H&N) and lung cohorts using clinically approved PBT plans. Main Results. All methods improved image quality compared to CBCT, with vCT achieving the best quantitative metrics (mean H&N MAE = 36.3HU; lung MAE = 39.9HU). However, image quality metrics did not correlate with dosimetric accuracy. In H&N cases, corrCBCT demonstrated inferior image metrics (MAE = 52.7HU) but achieved mean GPR of 98.4%, comparable to vCT and superior to DL-based methods. In lung cases, DDPM-generated sCTs achieved the highest mean GPR of 93.6% despite not having the best image metrics (MAE = 41.1HU). Dosimetric accuracy was highly sensitive to anatomical and Hounsfield Unit (HU) fidelity along beam paths, with beam angle significantly affecting GPR. DL-based methods effectively reduced artifacts but occasionally introduced anatomically inconsistent features that impacted dose calculation. Significance. Improvements in conventional image quality metrics did not necessarily translate to improved dosimetric accuracy in PBT. Instead, accurate anatomical representation and HU fidelity along beam paths were critical determinants of dose agreement. These findings highlighted the limitations of global image-based evaluation alone and underscored the need for dosimetric validation frameworks for sCT commissioning in PBT.