Synthetic CT (sCT) generation from cone-beam CT (CBCT) is a proposed enabler for adaptive radiotherapy workflows. The present study examined whether improvements in conventional image-quality metrics after CBCT-to-sCT conversion translate into equivalent improvements in dose calculation accuracy for proton beam therapy (PBT). The work specifically considered CBCTs acquired on proton gantries, which are prone to additional artifacts and noise, and compared deep-learning (DL) methods with non-learning commercial correction methods.
The authors evaluated two DL approaches for sCT generation: a denoising diffusion probabilistic model (DDPM) and a cycle-consistent generative adversarial network (GAN). These were compared against two non-learning-based correction methods available in a commercial treatment planning system: corrected CBCT (corrCBCT) and virtual CT (vCT).
Assessments were performed using head-and-neck (H&N) and lung patient cohorts with clinically approved PBT plans. Image quality was quantified using mean absolute error (MAE), peak signal-to-noise ratio, and normalized cross-correlation. Dosimetric accuracy was assessed using gamma passing rate (GPR) analysis and dose–volume histogram metrics, focusing on agreement between dose calculated on the sCT (or corrected CBCT/vCT) and the reference.
All evaluated methods improved global image quality relative to the original CBCT. Among approaches, vCT produced the best quantitative image-quality metrics overall: mean MAE for H&N cases was reported as 36.3 HU and for lung cases as 39.9 HU. Deep-learning methods and corrCBCT also reduced error compared with raw CBCT, but differed in the degree and pattern of improvement.
However, the study found that these conventional image-quality metrics were not predictive of dosimetric performance. Methods that achieved superior MAE or signal-to-noise improvements did not necessarily yield better dose agreement.
Dosimetric comparisons revealed divergent performance across methods and anatomical sites. In head-and-neck cases, corrCBCT yielded inferior image metrics (MAE = 52.7 HU) but achieved a mean GPR of 98.4%, which was comparable to vCT and superior to the DL-based methods in that cohort.
In lung cases, the DDPM-generated sCTs attained the highest mean GPR of 93.6% despite not having the best global image metrics (MAE = 41.1 HU). These findings demonstrate that global image metrics (for example, MAE) do not necessarily correlate with clinically relevant dose agreement as measured by GPR and dose–volume histogram comparisons.
The authors highlight that dosimetric accuracy in PBT was particularly sensitive to accurate anatomical representation and Hounsfield Unit (HU) fidelity along beam paths. Small inconsistencies in anatomy or HU values where proton beams traverse can substantially alter dose deposition due to the range sensitivity of protons.
Beam angle was also identified as a significant factor affecting GPR; dose agreement could vary with changes in beam path through regions where sCT methods differed in HU or anatomical portrayal. While DL-based approaches effectively reduced visible artifacts, they occasionally introduced anatomically inconsistent features that adversely affected dose calculation.
The primary implication is that improving global image-quality measures is insufficient as the sole criterion for adopting sCT methods into adaptive PBT workflows. Because proton dose calculations depend critically on accurate HU mapping and local anatomy along beam paths, the authors recommend that sCT evaluation for PBT include dedicated dosimetric validation rather than relying only on image-based metrics.
These results underscore the need for commissioning frameworks that incorporate dosimetric tests (for example, beam-path–specific checks, GPR analysis across relevant clinical beam angles, and dose–volume metric comparisons) before clinical deployment of sCT generation techniques in proton adaptive radiotherapy.
In CBCT-based adaptive proton radiotherapy, enhanced conventional image quality after CBCT-to-sCT conversion does not guarantee improved dosimetric agreement. Accurate anatomical representation and HU fidelity along beam paths, and sensitivity to beam angle, are key determinants of dose agreement. Deep-learning methods can reduce artifacts but may introduce localized anatomical inconsistencies that affect dose calculations. The study recommends dosimetric validation frameworks for sCT commissioning in PBT rather than reliance on global image-quality metrics alone.