Generative enhancement of suboptimally opacified cerebral CT angiography for large-vessel occlusion detection: a multireader study.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- Donghua University, Shanghai, China.
- Faculty of Medical Imaging Technology, College of Health Science and Technology, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
Abstract
Rapid identification of proximal anterior circulation large-vessel occlusion (LVO) on CT angiography (CTA) is essential for thrombectomy triage. Suboptimal arterial phase timing can obscure the intracranial arteries, prompting repeat CTA, additional contrast exposure, and treatment delay. We evaluated whether deep-learning-generated vascular images from a single suboptimally opacified acquisition improved LVO detection. This retrospective study included 626 patients. RescueCTA, a paired 2.5D conditional generative model, was developed in 391 patients using last-frame CT perfusion (CTP) as the simulated acquisition-failure input and same-session CTA as the target. Evaluation included 194 primary cases with simulated inputs and 41 exploratory cases with genuinely underopacified CTA. RescueCTA was compared with three generative baselines. Two radiologists assessed image quality. Four readers compared source images alone with source images plus RescueCTA for target LVO detection in the primary cohort; two also evaluated the exploratory cohort. The primary endpoint was the multireader multicase (MRMC) area under the receiver operating characteristic curve (AUC); exploratory analyses were descriptive. In the primary cohort, RescueCTA image quality was higher than that of all three baselines (3.90 ± 0.81). Combined reading improved AUC from 0.850 to 0.933 (ΔAUC = 0.083 [95% confidence interval (CI), 0.025-0.141]; <i>p</i> = 0.007), while paired CTA reached 0.948 (<i>p</i> = 0.594 vs. combined reading). Strict AUC improved from 0.756 to 0.880 (<i>p</i> < 0.001) but remained lower than that of paired CTA (0.935; <i>p</i> = 0.019). In the exploratory cohort, combined reading yielded a higher AUC than underopacified CTA alone (0.981 vs. 0.889). Deep-learning-generated vascular images improved LVO detection under simulated underopacification and showed preliminary feasibility in genuinely underopacified CTA. Larger real-world cohorts are needed to establish diagnostic performance after suboptimal arterial phase CTA.