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

INTRODUCTION

Stroke is a common cerebrovascular disease characterized by high incidence, mortality, disability, and recurrence rates.^[1] In recent years, it has exhibited a trend toward a younger onset.^[2] Early diagnosis and evaluation are therefore crucial for effective treatment and improved prognosis.^[3] Digital subtraction angiography (DSA) is considered the gold standard for vascular diagnosis and assessment owing to its high spatial resolution and real-time dynamic imaging capabilities.^[4] However, due to its risks, including high ionizing radiation exposure and the nephrotoxic effects of iodine-based contrast agents, DSA is generally not recommended as the first-choice imaging modality for patients with stroke.^[5]

Cerebral computed tomography perfusion (CTP) enables quantitative assessment of cerebral hemodynamics and accurate detection of abnormal cerebral blood perfusion in stroke patients.^[6,7] It can provide whole-brain perfusion maps and generate 4D computed tomography angiography (CTA) images. These images can replace traditional CTA for evaluating vasospasm while reflecting intracranial collateral circulation and vascular conditions.^[8-10] However, CTP requires continuous repeated scanning of the target area during contrast agent injection,^[6] which increases patients’ radiation exposure risk. Therefore, reducing radiation dose concerns is essential to ensure both patient safety and diagnostic effectiveness.

Beyond the effects of tube voltage, tube current, and scan frequency, previous studies have demonstrated that image reconstruction algorithms also significantly influence both radiation dose and image quality.^[11,12] With the rapid development of computer technology, iterative reconstruction (IR) algorithms have emerged as the mainstream reconstruction approach in computed tomography (CT) imaging. Adaptive statistical IR-Veo (ASIR-V; GE Healthcare, USA) is a vendor-specific hybrid IR technology. Compared with other IR methods, ASIR-V can reduce image noise and radiation dose while achieving shorter reconstruction times.^[13] However, similar to other IR techniques, the ASIR-V algorithm faces limitations, particularly the difficulty of balancing image noise reduction with maintaining natural image texture.^[14]

With the rapid advancement of artificial intelligence, deep learning technology has been widely applied to CT image reconstruction. Deep learning image reconstruction (DLIR) algorithms have been demonstrated to reduce noise and enhance diagnostic accuracy under low-dose conditions.^[15] By training on large-scale, high-quality conventional-dose CT images and their corresponding low-dose images or raw original projection data, DLIR algorithms learn to distinguish between true signals and noise, capturing both local detail and global image structure.^[16] This technology has been applied in various clinical settings, including pediatric “dual-low” chest CTA and abdominal dual-energy CT.^[17,18] By modeling the natural noise distribution of normal tissues, DLIR can substantially reduce radiation dose while preserving the authenticity of noise texture, maintaining high spatial resolution, and clearly depicting critical anatomical details such as fine structures and small vessel branches,^[19] thereby producing higher-quality images. However, the clinical value of DLIR for improving cranial CTA image quality based on CTP data has not been fully explored.

Therefore, this study was performed to evaluate the image quality of CTA using DLIR in CTP compared with the ASIR-V algorithm. In addition, it assessed the diagnostic performance of these reconstruction methods against DSA in detecting arterial stenosis.

MATERIAL AND METHODS

The Medical Ethics Committee of the Affiliated Hospital of Xuzhou Medical University approved this study (XYFY2025-KL284-01). In total, 65 patients who underwent CTP examinations at our hospital between June 2024 and June 2025 were retrospectively enrolled. Of these, 14 patients underwent both CTP and DSA examinations [Figure 1].

Close

Figure 1:

Flowchart of this study. (ASIR-V: adaptive statistical iterative reconstruction-veo; DLIR-L: deep learning low setting; DLIR-H: deep learning high setting).

Export to PPT

Close

Figure 2:

Schematic illustration of edge rise distance (ERD) and edge rise slope (ERS) measurements. (a) A straight line perpendicular to the wall of the basilar artery is drawn. (b) Red dots mark the 10% and 90% points. ERD is defined as the distance between these two points on the computed tomography attenuation curve, while ERS is the slope between them.

Export to PPT

RESULTS

Post hoc comparison of objective image quality evaluation

Pairwise comparisons showed no significant differences in CT values among the four reconstruction algorithms: ASIR-V40%, ASIR-V80%, DLIR-L, and DLIR-H (G1–G4) for any vascular region (p > 0.05). In the ICA and BA, SD values in the G4 group were significantly lower than those of the G1 and G2 groups (p < 0.05). Regarding both CNR and SNR, no significant difference was observed between the G2 and G4 groups (p > 0.05); however, both groups exhibited significantly higher CNR and SNR values than the G1 and G3 groups (p < 0.05) [Table 5, Figure 3].

Table 5: Post hoc comparison of objective image quality evaluation.

Measurement metrics	G1 versus G2	G1 versus G3	G1 versus G4	G2 versus G3	G2 versus G4	G3 versus G4
MCA
CT value	0.987	0.356	0.442	0.348	0.433	0.877
SD	0.554	0.264	0.008	0.573	0.029	0.094
CNR	<0.001	0.579	<0.001	0.004	0.265	<0.001
SNR	<0.001	0.773	0.005	0.002	0.427	0.016
ICA
CT value	0.980	0.503	0.657	0.482	0.652	0.760
SD	0.160	0.049	<0.001	0.572	0.003	0.019
CNR	<0.001	0.942	<0.001	<0.001	0.416	<0.001
SNR	<0.001	0.970	0.003	<0.001	0.328	0.005
BA
CT value	0.612	0.092	0.213	0.247	0.002	0.931
SD	0.612	0.153	0.001	0.349	0.005	0.048
CNR	<0.001	0.524	<0.001	0.007	0.131	<0.001
SNR	<0.001	0.539	0.003	0.003	0.684	0.017
Temporalis
CT value	0.703	0.618	0.339	0.830	0.469	0.640
SD	<0.001	0.673	<0.001	<0.001	0.470	<0.001

G1-G4 as defined in the main text. Bold type indicates significant differences. MCA: Middle cerebral artery, CT: Computed tomography, SD: Standard deviation, CNR: Contrast-to-noise ratio, SNR: Signal-to-noise ratio, ICA: Internal carotid artery, BA: Basilar artery, Temporalis: Temporalis muscle

Close

Figure 3:

Post hoc comparison of objective image quality evaluation (contrast-to-noise ratio, signal-to-noise ratio). (a) Violin plot of middle cerebral artery (MCA) contrast-to-noise ratio (CNR). (b) Violin plot of MCA signal-to-noise ratio (SNR). (c) Violin plot of internal carotid artery (ICA) CNR. (d) Violin plot of ICA SNR. (e) Violin plot of basilar artery (BA) CNR. (f) Violin plot of BA SNR.

Export to PPT

Post hoc comparison of subjective evaluation indicators

Subjective scoring results showed that G4 had the highest proportion of scores of >4 points for image noise, overall image quality, and vascular edge sharpness [Figure 4]. Post hoc comparisons revealed that G4 was significantly superior to G1 and G2 for all three indicators (p < 0.001). Compared with G3, G4 also showed significantly higher scores for image noise and overall image quality but no significant difference in vascular edge sharpness (p = 0.60) [Table 7].

Close

Figure 4:

Stacked bar chart of subjective image quality scores (5 points = excellent). (a) Distribution of image noise. (b) Distribution of vascular edge sharpness. (c) Distribution of overall image quality. Light blue: >4 (Excellent); Blue: >3 & ≤4 (Good); Orange: ≤3 (Average).

Export to PPT

Table 7: Post hoc comparison of subjective evaluation indicators.

Measurement metrics	G1 versus G2	G1 versus G3	G1 versus G4	G2 versus G3	G2 versus G4	G3 versus G4
Image noise	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001
Vascular edge sharpness	0.044	<0.001	<0.001	<0.001	<0.001	0.6
Overall image quality	0.117	0.147	<0.001	0.925	<0.001	<0.001

G1-G4: Corresponding to four distinct image reconstruction schemes described in the main text (specific scheme descriptions align with the main body content). Bold values indicates statistically significant differences (P<0.05) between the compared groups.

Evaluation of arterial stenosis degree

DSA identified a total of 26 vascular stenoses: 1 case of mild stenosis, 1 case of moderate stenosis, 5 cases of severe stenosis, and 19 cases of total occlusion. Statistical analysis showed that the DLIR group had the highest agreement with DSA (κ = 0.819, p < 0.001), whereas agreement between the ASIR-V group and DSA was only moderate (κ = 0.422, p = 0.008) [Table 8]. Examples of arterial stenosis cases have been shown in Figure 5.

Close

Figure 5:

Severe stenosis of the M1 segment of the right middle cerebral artery (MCA) on digital subtraction angiography (DSA) and computed tomography angiography (CTA) reconstructions using different algorithms. A 57-year-old male underwent both computed tomography perfusion and DSA examinations. (a) The DSA image shows severe stenosis in the M1 segment of the right MCA (white arrow). Reconstructed axial MIP images from the CTA using (b) adaptive statistical iterative reconstruction-veo (ASIR-V) 40%, (c) ASIR-V 80%, (d) deep learning image reconstruction (DLIR)-L, and (e) DLIR-H also demonstrate severe stenosis in the same segment (white arrow). Figure (a), DSA image: yellow box = stenotic right MCA M1 segment; Figures (b) to (e), CTA reconstruction images: yellow box = corresponding DSA stenotic segment. The yellow boxes in (b)–(e) indicate the corresponding DSA stenotic segment on the CTA reconstructions.

Export to PPT

Table 8: Evaluation of arterial stenosis degree.

DSA (n=26)	ASIR-V 40% (n=26)				ASIR-V 80% (n=26)				DLIR-L (n=26)				DLIR-H (n=26)
	M	Mod	Sev	Occl	M	Mod	Sev	Occl	M	Mod	Ser	Occl	M	Mod	Sev	Occl
M	1	0	0	0	1	0	0	0	1	0	0	0	1	0	0	0
Mod	0	0	1	0	0	0	1	0	0	1	0	0	0	1	0	0
Sev	0	0	2	3	0	0	2	3	0	0	4	1	0	0	4	1
Occl	0	0	2	17	0	0	2	17	0	0	1	18	0	0	1	18
P-value	0.008				0.008				<0.001				<0.001
κ-value	0.422				0.422				0.819				0.819

M: Mild stenosis, Mod: Moderate stenosis, Sev: Severe stenosis, Occl: Total occlusion, DSA: Digital subtraction angiography, ASIR-V: Adaptive statistical iterative reconstruction-Veo, DLIR: Deep learning image reconstruction. Vessels that appeared normal on the gold standard DSA images were not included in the comparative evaluation. Bold type indicates significant differences. Bold values: Bold values indicates statistically significant differences (P<0.05) between the compared groups.

DISCUSSION

In this study, we proposed a more effective reconstruction protocol for cerebral CTA based on CTP data under a low radiation dose. Our results showed that the DLIR-H algorithm was significantly superior to both ASIR-V and DLIR-L in terms of image noise control, CNR, SNR, and diagnostic consistency for vascular stenosis.

Prior studies have compared the performance of ASIR-V and DLIR in different CT applications.^[22,23] In our study, DLIR and ASIR-V were compared specifically for reconstructing cerebral CTA images from 80kVp CTP data. DLIR-H demonstrated markedly better noise suppression than both ASIR-V and DLIR-L at all vascular levels. For example, compared with ASIR-V 40%, DLIR-H reduced the SD by approximately 15.1% (25.2 vs. 21.4), 28.6% (23.8 vs. 17.0), and 16.9% (25.4 vs. 21.1) in the MCA, ICA, and BA. Even against ASIR-V 80%, which has high iterative intensity, DLIR-H still reduced ICA noise by about 22.0% (21.8 vs. 17.0, p < 0.001), demonstrating greater noise suppression efficiency. These findings are consistent with those of Jiang et al.,^[24] who reported that in ultra-low-dose chest CT for pulmonary nodule detection, DLIR significantly reduced image noise and improved nodule detection rates compared with ASIR-V.

Regarding CNR and SNR, there was no significant difference between the DLIR-H and ASIR-V 80% groups, but both groups were significantly superior to the ASIR-V 40% and DLIR-L groups (p < 0.001). This noise reduction pattern may stem from the convergence of the two reconstruction modalities’ respective technical advantages and inherent performance ceilings: ASIR-V 80% operates at high iterative intensity, where its noise suppression capacity approaches the upper limit of the statistical iterative framework. Through repeated projection data optimization, it effectively mitigates image noise. In contrast, DLIR-H leverages a deep learning architecture trained on massive imaging datasets to discriminate noise from valid signals; thus, within the specific context of cerebrovascular imaging in the present study, the noise reduction efficacy of these two fundamentally distinct mechanisms achieved a statistically equivalent level of performance. Lei et al.^[25] reported in their image quality assessment study that increasing DLIR reconstruction intensity significantly improved both CNR and SNR in white matter and gray matter. This observation highlights the superior utility of DLIR in elevating contrast and signal fidelity for brain tissue imaging. The core conclusion of their work aligns with ours: both studies confirmed that DLIR-H achieves the highest CNR and SNR, with performance that markedly outperforms low-intensity IR. However, Lei et al.^[25] centered their analysis on white matter and gray matter, analyzing the effects of DLIR on brain tissue density differences and perfusion parameters. In contrast, our study targeted cerebrovascular structures (MCA, ICA, and BA). By measuring density differences between vessels and the temporalis muscle to calculate CNR and SNR, we specifically verified the value of DLIR in improving contrast resolution in complex vascular imaging, providing complementary evidence for cerebral CTA vascular assessment based on CTP data.

Previous studies have already confirmed that DLIR-H achieved the highest scores in subjective evaluation of vessels.^[26-27] In our study, DLIR-H achieved the highest ratings for image noise, overall image quality, and vascular edge sharpness and was significantly superior to all other groups. This discrepancy in subjective performance stems from the inherent limitations of statistical IR versus the advantages of deep learning-based frameworks: while ASIR-V80% enhances noise reduction by increasing iterative intensity, its reliance on mathematical model-driven projection data optimization tends to cause excessive smoothing.^[28] This smoothing effect compromises fine anatomical detail preservation and may induce “waxy” artifacts, particularly in complex vascular regions.^[18,25,28] In contrast, DLIR-H is built on convolutional neural networks trained on large-scale high-dose imaging datasets, enabling it to selectively suppress noise while retaining native image texture and vascular edge details.^[29] This unique advantage allows DLIR-H to outperform ASIR-V80% in subjective quality, addressing the trade-off between noise reduction and detail. Our findings demonstrate that DLIR-H effectively mitigates high image noise associated with low-dose 80 kVp scanning, which arises from insufficient photon flux,^[30-31] while facilitating the precise visualization of vascular anatomical details and ensuring patient radiation safety. To further validate the subjective evaluation of vascular edge sharpness, we measured ERD and ERS of the BA as objective indicators. Both ERD and ERS were significantly better in the DLIR groups than in the ASIR-V groups (P < 0.05), with DLIR-H performing best overall, owing to DLIR-H images exhibiting significantly increased CT attenuation at vascular-tissue interfaces, which contributed to substantially higher ERS values. These findings confirm that DLIR-H can improve vascular edge sharpness in cerebral CTA reconstructed from CTP data.

In addition, unlike previous studies on brain CTA reconstructed from CTP, this study used DSA as a reference to compare the accuracy of four reconstruction techniques in diagnosing arterial stenosis on CTA. Our results showed that DLIR demonstrated significantly higher agreement with DSA in stenosis assessment (κ = 0.819, p < 0.001) compared with ASIR-V (κ = 0.422, p = 0.008). This finding is consistent with Qu et al.,^[32] who reported that the DLIR’s diagnostic accuracy in lower extremity CTA approaches that of DSA. Previous studies have revealed that DLIR optimizes reconstruction through a deep learning model, markedly improving vascular boundary clarity while suppressing noise. Qu et al.^[32] showed that in low-dose lower extremity CTA, the ERS of DLIR-H was significantly higher than that of ASIR-V, indicating clearer depiction of vascular wall details. Wang et al.^[33] demonstrated in coronary CTA that DLIR-H provides higher spatial resolution for visualizing calcified plaques and intra-stent lumens, offering key morphological evidence for stenosis diagnosis. Compared to ASIR-V, the model characteristics of DLIR make it more suitable for evaluating vascular details under low-contrast conditions. In a phantom-based study, Euler et al.^[34] found that for low-contrast tissues, especially at low radiation doses, the spatial resolution of ASIR-V was inferior to that of FBP, with the effect worsening at higher ASIR-V blending levels. By contrast, DLIR-H facilitates the optimization of image noise distribution: It significantly suppresses image noise while enhancing spatial resolution and preserving native image texture, ultimately resulting in clearer vascular boundaries.^[35] These advantages likely explain the higher consistency of DLIR with DSA in stenosis diagnosis observed under the specific conditions of 80 kVp, low radiation, and low contrast in our study.

However, this study has several limitations. First, as a single-center retrospective study with a relatively small sample size, its findings require further validation through larger, multicenter datasets. Second, the range of reconstruction intensities compared was limited. Only ASIR-V 40%, ASIR-V 80%, and DLIR were included; other iterative intensity levels were not assessed, making it difficult to fully characterize the impact of different reconstruction parameters on image quality. Finally, the use of the gold standard was limited because relatively few patients underwent concurrent DSA examinations, which may reduce the strength of our verification of CTA-based stenosis diagnosis.

CONCLUSION

The DLIR-H algorithm significantly improves the image quality of cerebral CTA reconstructed from CTP data by effectively reducing noise, enhancing CNR and SNR, and optimizing diagnostic performance. It can therefore be recommended as the preferred clinical reconstruction approach.