Abstract
Background/Objectives: Virtual contrast-enhanced CT (vCECT) synthesis may support imaging workflows when iodinated contrast administration is undesirable. However, whole-image similarity can obscure errors in localized vascular enhancement. We evaluated a mask-guided two-stage framework that couples spatially targeted enhancement with separate assessment of localization and synthesis. Methods: This retrospective study included 50 patients with 2550 paired thoracic non-enhanced CT (NECT) and contrast-enhanced CT (CECT) slices from radiotherapy simulation. Stage 1 predicted a contrast-region mask from NECT; Stage 2 used this mask to guide residual synthesis, trained with locally blended pseudo-contrast targets. Localization was assessed using segmentation metrics, and synthesis was evaluated against reference CECT over the whole image and reference-mask regions of interest (ROIs). Ablations compared single-stage residual synthesis, predicted-mask guidance, and reference-mask guidance. Results: Mask prediction achieved a mean Dice similarity coefficient (DSC) of 0.91. Whole-image peak signal-to-noise ratio (PSNR) and structural similarity index were 46.91 dB and 0.96, respectively; ROI values were 34.22 dB and 0.95. Compared with single-stage synthesis, predicted-mask guidance increased ROI PSNR from 18.97 to 34.22 dB and reduced normalized root mean squared error from 0.12 to 0.03. Reference-mask guidance achieved an ROI PSNR of 36.13 dB. Conclusions: Explicit localization improved regional synthesis fidelity in this cohort. The framework links localized CT enhancement to quantitative image-quality evaluation; external and reader-based validation remain necessary before clinical use.