Optimization of image quality and diagnostic accuracy for vascular stenosis in computed tomography perfusion-reconstructed cerebral computed tomography angiography using deep learning reconstruction

Objectives: This study aims to compare the effects of deep learning image reconstruction (DLIR) and adaptive statistical iterative reconstruction-veo (ASIR-V) algorithms on the image quality of cranial computed tomography angiography (CTA) reconstructed from computed tomography perfusion (CTP) data and to evaluate their agreement with digital subtraction angiography (DSA), the gold standard for diagnosing arterial stenosis. Material and Methods: This retrospective study enrolled patients with clinically suspected cerebrovascular disease who had undergone CTP examination. From the arterial peak phase of CTP raw data, four CTA datasets were reconstructed: ASIR-V 40%, ASIR-V 80%, DLIR low setting (DLIR-L), and DLIR high setting (DLIR-H). Regions of interest were placed at the M1 segment of the middle cerebral artery (MCA), basilar artery (BA), and internal carotid artery (ICA) on the healthy side and the temporalis muscle. Computed tomography attenuation values and standard deviations were measured, and the contrast-to-noise ratio(CNR) and signal-to-noise ratio (SNR) were subsequently calculated using the attenuation values and standard deviations. Vessel edge sharpness was objectively assessed using edge rise distance (ERD) and edge rise slope (ERS). Using DSA as the reference standard, diagnostic performance for vascular stenosis was evaluated, and agreement was analyzed with the Kappa test. Two radiologists independently scored subjective image quality using a 5-point Likert scale. Results: Sixty-five patients were included, 14 of whom additionally underwent DSA. For the MCA, ICA, and BA, image noise in the DLIR-H group was significantly lower than that in the other groups, with statistically significant differences compared with ASIR-V 40% ( p < 0.05). For CNR and SNR, DLIR-H and ASIR-V 80% outperformed ASIR-V 40% and DLIR-L across all arterial levels ( p < 0.05). ERD and ERS were significantly superior in the DLIR groups compared with the ASIR-V groups ( p < 0.05). In subjective evaluations, the DLIR-H group achieved the highest scores across all parameters ( p < 0.05). Agreement with DSA was excellent for DLIR-H (κ = 0.819), significantly higher than for ASIR-V ( p < 0.05). Conclusion: In CTP examinations for suspected cerebrovascular disease, DLIR-H markedly improves cranial CTA image quality compared with ASIR-V, while achieving higher diagnostic agreement with DSA. DLIR-H can be recommended as the preferred clinical reconstruction method.