Author: Yizheng Chen, Michael Gensheimer, Mingjie Li, Lei Xing 👨🔬
Affiliation: Department of Radiation Oncology, Stanford University 🌍
Purpose: Automatically translating non-contrast to contrast-enhanced computed tomography (CT) images is critical for improving clinical workflow, reducing heathcare cost, minimizing radiation exposure, and preventing contrast agent allergies. Toward the goal of providing such a solution that enhance vascular structures without altering underlying anatomical details, we develop a novel conditional diffusion model guided by text prompts, offering a flexible and generalizable approach to contrast CT generation.
Methods: We leverage a pretrained stable diffusion framework that denoises non-contrast CT images in the latent space to establish a robust model for the generation of their contrast-enhanced counterparts. Text prompts were used as explicit instructions to control the desired phase of contrast enhancement (e.g., arterial or venous). The model was trained on a multi-center, multi-phase, and multi-bodypart dataset, including over 70,000 paired images from our Hospital and the publicly available Coltea-Lung-CT-100W dataset. Performance was evaluated using structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) metrics, compared against state-of-the-art (SOTA) methods.
Results: The proposed method demonstrated superior performance, achieving the highest SSIM and PSNR values compared to existing techniques for non-contrast to contrast CT translation. The model’s generalizability was validated across diverse clinical scenarios, with its ability to handle multi-body part and multi-phase data and adapt the contrast phase dynamically through text prompts.
Conclusion: This study introduces a novel diffusion-based framework for generating contrast-enhanced CT images with high accuracy and flexibility. By offering a unified and efficient solution, the approach has significant potential for clinical adoption, paving the way for safer and more cost-effective CT imaging workflows.