AI-based Synthetic CT Angiography from Non-contrast CT for Cerebrovascular Disease Screening.
Authors
Affiliations (9)
Affiliations (9)
- Department of Neurosurgery, the 904th Hospital of Joint Logistic Support Force of PLA, Wuxi Clinical College of Anhui Medical University, Wuxi 214044, Jiangsu, China.
- Department of Radiology, Changhai Hospital, Naval Medical University, Shanghai 200433, China.
- Department of Pharmacy, the 904th Hospital of Joint Logistic Support Force of PLA, Wuxi 214044, Jiangsu, China.
- Department of Neurosurgery, Wuxi School of Medicine, Jiangnan University, Wuxi 214122, Jiangsu, China.
- Department of Neurosurgery, Wuxi People's Hospital Affiliated to Nanjing Medical University, Wuxi 214023, Jiangsu, China.
- Department of Neurosurgery, Air Force Medical Center, Air Force Medical University, PLA, Beijing 100142, China.
- School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214122, Jiangsu, China.
- Department of Neurosurgery, Shanghai Changhai Hospital, Naval Medical University, Shanghai 200433, China.
- Department of Neurosurgery, Beijing Tiantan Hospital, Capital Medical University, Beijing 100070, China.
Abstract
CT angiography (CTA) is a key investigation in cerebrovascular disease. However, CTA is not always available and it also requires intravenous injection of iodinated contrast agents. It is increasingly possible to derive additional information from standard imaging sequences using Artificial Intelligence techniques. We investigated whether CTA maps could be derived from non-contrast CT (NCCT). We conducted a retrospective, multicenter study across five Chinese hospitals, enrolling 3,709 patients who underwent head NCCT paired with CTA. The dataset encompassed three cerebrovascular conditions: intracranial aneurysms (IA), intracranial atherosclerotic stenosis (IAS), and normal intracranial arteries. We developed the Cerebrovascular CTA Generative Artificial Intelligence Model (CTA-GAI) to synthesize CTA images directly from NCCT head scans. Synthetic outputs were evaluated against five typical models using quantitative metrics and visual assessment by clinicians. We also assessed the potential clinical utility of synthetic CTA for preliminary screening and triage by evaluating its ability to distinguish diseased from normal intracranial arteries and to classify common cerebrovascular subtypes. CTA-GAI demonstrated consistent and robust performance across both the validation and test sets. In internal validation, synthetic CTA images achieved a mean absolute error (MAE) of 0.0416, mean squared error (MSE) of 0.0178, peak signal-to-noise ratio (PSNR) of 25.59 dB, and structural similarity index measure (SSIM) of 85.41%. Clinicians assigned an average visual quality score of 4.45 out of 5. These metrics reflect close approximation to real CTA images. Performance remained consistent across four external validation sets. In a test set of 110 patients, clinicians achieved an overall accuracy, precision, sensitivity, specificity, and F1 score of 92.7%, 97.7%, 86.0%, 98.3%, and 91.5%, in distinguishing diseased intracranial arteries. Differentiation between IA and IAS within diseased arteries reached 90.7% accuracy. CTA-GAI can synthesize CTA-like images from NCCT that show promising utility for preliminary assessment in clinical practice. These results support its potential role as a rapid, low-cost, and non-invasive tool for large-scale screening or triage of cerebrovascular diseases.