基于平扫CT生成增强CT的技术与临床应用进展OA

Contrast-enhanced computed tomography(CECT)uses iodinated contrast agents to improve the contrast be-tween the lesion and adjacent tissue,and it plays a critical role in tumor staging,evaluation of vascular diseases,emergency screening,radiation therapy planning,and preoperative planning;however,its clinical application is limited by allergy-like reac-tions,contrast-associated acute kidney injury,complicated workflows,increased medical costs,and prolonged emergency manage-ment time.Non-contrast computed tomography(NCCT)is safe and convenient,and deep learning models have shown the promis-ing performance of NCCT in the screening of pancreatic cancer and colorectal cancer and the assessment of acute ischemic stroke;nevertheless,insufficient soft tissue contrast still limits the detection of hypovascular lesions and small lesions.Generation of vir-tual CECT images from NCCT based on artificial intelligence aims to retrieve enhancement information of blood vessels,parenchy-mal organs,lymph nodes,and tumors without administering additional contrast agents,thereby reducing the risk of patients and optimizing the allocation of medical resources.This article summarizes the mainstream technical approaches including generative adversarial network(GAN),conditional GAN,CycleGAN,Wasserstein GAN,multi-stage cascaded networks,CNN-Transformer hybrid architectures,diffusion probabilistic models,and conditional diffusion models,as well as the application of anatomical pri-ors,segmentation labels,organ/vessel masks,multi-task learning,clinical parameters,and multimodal information.Related stud-ies have been extended to multiple scenarios such as CTA generation,multi-phase hepatic contrast-enhanced imaging,abdominal tumor screening,identification of large-vessel occlusion in acute ischemic stroke,and segmentation of organs at risk for radio-therapy.Overall,although this technique holds great potential for reducing the use of contrast agents,assisting clinical diagnosis,and optimizing clinical workflows,it still faces several challenges including a lack of paired datasets,limited cross-center general-ization capability,imaging artifacts and model hallucinations,as well as undefined safety thresholds and regulatory frameworks in clinical practice.