Synthetic Contrast Enhancement from Non‑Contrast CT for Improved Vascular Visualization In the Pediatric Abdomen Using a Sequential Slab-Based 3D Adversarial Diffusion Model
Abstract
Purpose
Accurate vascular visualization is essential for target delineation in pediatric radiotherapy, particularly for abdominal tumors such as Wilms tumor and neuroblastoma, where major vessels serve as key anatomical landmarks. Although contrast-enhanced CT (CECT) provides optimal depiction, iodinated contrast use and prolonged anesthesia may increase procedural complexity. We present a slab-based 3D diffusion model that synthesizes contrast enhancement from non-contrast CT (NCCT) to improve visualization of target-relevant vascular structures.
Methods
An image translation model was developed using a 3D denoising diffusion framework with an adversarial diffusion backbone. The network leverages adjacent slices with a four‑slice interval to capture inter‑slice continuity while reducing model complexity and accommodating variable slice counts typical of abdominal CT. The model is bidirectional, enabling synthesis of CT (SCECT) from NCCT and virtual NCCT from CECT. Performance was evaluated against a 2D adversarial diffusion image‑to‑image translation method using pediatric abdominal CT datasets (training n = 163, testing n = 10).
Results
Vascular regions of interest (ROIs), including major abdominal vessels and primary branches, were generated using subtraction images between CECT and NCCT, followed by manual refinement and a 2‑mm dilation to include surrounding tissue. The model significantly outperformed the 2D approach across abdominal vascular soft‑tissue ROIs (20–200 Hounsfield Unit, HU), demonstrating higher PSNR (40.7 ± 3.4 vs. 27.6 ± 2.1), and lower RMSE (3.4 ± 1.6 vs. 10.1 ± 4.7). Qualitative evaluation showed sharper vascular boundaries and improved slice‑to‑slice consistency. HU analysis within vascular ROIs demonstrated close agreement between SCECT and CECT, with mean values of 87.3 ± 31.0 and 93.1 ± 31.9, respectively.
Conclusion
The proposed adversarial diffusion approach preserves vascular anatomy and HU characteristics while synthesizing CECT from NCCT. This framework offers a practical and efficient strategy for generating vascular-rich synthetic images when CECT is limited or undesirable.