Commercial AI platforms improve reproducibility but fail to meet clinical precision thresholds for intracranial aneurysm measurement: implications for serial surveillance

Abstract

Background

Intracranial aneurysm affect 3–7% of the global population, with rupture causing > 80% of non-traumatic subarachnoid hemorrhage and approximately 50% mortality. Clinical management relies on precise measurement of aneurysm neck width and maximum length, where ≥ 1 mm growth signals elevated rupture risk. computed tomography angiography enables non-invasive monitoring but manual measurements suffer from inter-observer variability. Commercial artificial intelligence platforms offer potential improvements, yet their consistency and accuracy versus digital subtraction angiography, the gold standard, are understudied.

Methods

This retrospective study analyzed 148 patients with 163 Intracranial aneurysms via computed tomography angiography, including a subgroup of 86 with digital subtraction angiography within 1 week. Measurements were obtained using Shukun artificial intelligence, UIH artificial intelligence, manual computed tomography angiography (intra-observer repeated at 1 month), and digital subtraction angiography. Reproducibility was assessed by coefficient of variation; agreement by Bland-Altman analysis, with 95% limits of agreement within ± 1.0 mm considered clinically acceptable.

Results

Results For aneurysm neck width, manual measurements had a mean difference of -1.62 mm and 95% limits of agreement of -4.87 to 1.62 mm vs. digital subtraction angiography, while UIH artificial intelligence and Shukun artificial intelligence had mean differences of + 0.80 mm and − 1.01 mm, and 95% limits of agreement of -3.68 to 2.08 mm and − 3.06 to 1.04 mm, respectively. For aneurysm maximum length, UIH artificial intelligence systematically overestimated (mean difference: +3.46 mm) and Shukun artificial intelligence underestimated (mean difference: -2.20 mm) vs. digital subtraction angiography. Both artificial intelligence platforms had narrower coefficients of variation (0.31–0.36 for aneurysm neck width, 0.23–0.26 for aneurysm maximum length) than manual measurements (0.36–0.42, 0.23–0.24). However, all methods exceeded the clinically acceptable ± 1.0 mm threshold.

Conclusions

Artificial intelligence platforms have better reproducibility than manual measurements but show systematic biases, not meeting the clinical ± 1.0 mm precision vs. the digital subtraction angiography gold standard. Inter-platform variability exceeds the aneurysm growth threshold, requiring consistent use of the same artificial intelligence platform in serial surveillance to prevent misinterpreting changes.

Introduction

Intracranial aneurysm (IA) affect an estimated 3–7% of the global population [1], and their rupture accounts for > 80% of non-traumatic subarachnoid haemorrhage [2], with mortality approaching 50% [3].The clinical management of IA, ranging from conservative monitoring to invasive endovascular or neurosurgical intervention, hinges entirely on the accurate assessment of key morphological parameters, particularly aneurysm neck width and maximum diameter [4]. A size increase of ≥ 1 mm between serial imaging studies is widely recognized as a surrogate marker for elevated rupture risk, making precise and consistent measurement not only a diagnostic cornerstone but also a determinant of treatment decision-making and long-term follow-up strategies [5].

Digital subtraction angiography (DSA) has long been regarded as the gold standard for IA measurement due to its high spatial resolution and ability to visualize vascular anatomy in dynamic detail [6]. However, DSA is an invasive procedure associated with risks such as contrast-induced nephropathy and vascular complications, limiting its utility for routine screening or repeated follow-up scans [7]. In contrast, computed tomography angiography (CTA) offers a non-invasive alternative with faster acquisition times and broader accessibility, making it the preferred modality for initial IA detection and serial monitoring in clinical practice [8].

However, manual CTA mensuration is labour-intensive, operator-dependent, and prone to inter-observer variability [9]. In recent years, artificial intelligence (AI) technology has advanced rapidly in medical imaging, with deep learning algorithms enabling automated detection, segmentation, and morphological parameter measurement of IA, significantly improving diagnostic efficiency and accuracy [10]. Numerous studies have validated the accuracy and speed of AI in IA detection, garnering broad recognition. Yet, research on the consistency and accuracy of AI for measuring IA-related parameters remains relatively scarce [11]. Critical gaps persist: few studies have systematically compared different commercial AI platforms against each other, and even fewer have rigorously evaluated their accuracy relative to the DSA gold standard [9]. Additionally, the extent to which AI outperforms manual measurement in stability (e.g., intra-reader consistency) and whether AI-derived CTA measurements can reliably substitute for DSA in treatment planning remain unclear-gaps that hinder the safe integration of AI into routine IA management.

To address these gaps, the present study aimed to comprehensively evaluate the accuracy and consistency of IA measurement using two commercial AI software platforms (Shukun and United Imaging Healthcare [UIH]) and manual methods, with DSA as the reference standard. Specifically, our objectives were: (1) to quantify the intra-rater variability of manual CTA measurements (assessed by the same radiologist at a 1-month interval) and inter-method variability between AI and manual CTA measurements using the coefficient of variation (CV); (2) to assess the consistency between the two AI platforms (Shukun vs. UIH) and between each AI platform and DSA using Bland-Altman analysis; (3) to compare the magnitude of bias between AI-derived and manual CTA measurements relative to DSA, and to determine whether either method meets the clinical requirement of a consistency limit within the ± 1 mm threshold for IA growth.

Materials and methods

Study population and eligibility criteria

This single-center retrospective study was conducted at Zhuzhou Central Hospital (Institutional Review Board[IRB]-approved; informed consent waived due to retrospective design). Consecutive patients who underwent CTA for IA diagnosis between August 2022 and February 2025 were screened, among whom a subset also underwent DSA.

Inclusion criteria: (1) IA confirmed at the same anatomical site by ≥ 2 modality (Shukun AI, UIH AI, or DSA); (2) CTA-DSA interval ≤ 1 week; (3) Age ≥ 18 years.

Exclusion criteria: (1) Prior surgical clipping of target IA; (2) CTA/DSA with motion/technical artifacts precluding measurement; (3) Concurrent intracranial pathologies (e.g., tumors, arteriovenous malformations[AVMs], prior cranial surgery); (4) Incomplete data.

A final cohort of 148 patients harboring 163 IA was included for the CTA-based analysis. From this cohort, a subgroup of 86 patients who had additionally undergone DSA formed the basis for the comparative analysis against the reference standard(Fig. 1).

Fig. 1.

Study design and workflow. Schematic overview of the study design showing patient enrollment, measurement methods and the analytical workflow. CTA, computed tomography angiography; DSA, digital subtraction angiography; AI, artificial intelligence

CTA and DSA protocols

CTA examination

CTA was performed using 3 clinically validated scanners: dual-source SOMATOM Force, 64-slice SOMATOM Definition AS (Siemens, Munich, Germany), and 80-slice uCT 780 (UIH, Shanghai, China). Uniform acquisition parameters were applied: tube voltage 120 kV, tube current 300 mA (with automatic tube current modulation), pitch 0.6, rotation time 0.5 s, collimation 64 × 0.6 mm (for 64-slice scanner) or 80 × 0.5 mm (for 80-slice scanner), slice thickness 0.75 mm with 0.5 mm reconstruction interval, and standard kernel reconstruction.

Non-ionic iodinated contrast (iohexol 350 mgI/mL or iopamidol 370 mgI/mL) was injected intravenously (4.0–5.0 mL/s, 60–80 mL), followed by 30 mL saline flush. Scanning was triggered via bolus tracking (100 Hounsfield Unit [HU] threshold, ascending aorta).

DSA examination

DSA was performed in an Azurion suite (Philips, Amsterdam, Netherlands) for endovascular candidates or ambiguous CTA findings. Under local anesthesia, a 5-Fr sheath was inserted via femoral puncture; 4-Fr catheters selectively cannulated the internal carotid artery and vertebral artery. Iopromide (370 mgI/mL) was injected (3–5 mL/s, 5–7 mL/injection). Multiple projections (anteroposterior, lateral, 30°–45° oblique) were acquired at 1024 × 1024 pixels.

Measurement methods

All imaging data were anonymized per Health Insurance Portability and Accountability Act (patient identifiers replaced with unique codes). Two key IA parameters were measured:

• Aneurysm neck width: Maximum transverse diameter at the neck-parent vessel junction.

• Maximum length: Longest linear distance from neck to aneurysm dome.

Manual measurement

CTA measurements: Performed on GE Picture Archiving and Communication System (Version 4.0) by a neuroradiologist with 8 years of cerebrovascular experience, using multiplanar reconstruction, curved planar reconstruction, and volume rendering. Intra-observer consistency was assessed via repeated measurements (1-month interval, blinded to initial results).

DSA measurements: Conducted on a dedicated workstation (rotational 3D DSA) by two board-certified neuroradiologists (8/10 years of experience). Mean values were used; discrepancies > 1.0 mm were resolved via joint review.

AI-based measurement

CTA Digital Imaging and Communications in Medicine files were uploaded to two AI platforms: Shukun AI (Version 1.0.8, Shukun Tech, Beijing) and UIH AI (Version R001, United Imaging Healthcare, Shanghai). Both AI platforms were pre-trained on proprietary datasets (manufacturer-disclosed training set sizes: Shukun AI > 5,000 IA, UIH AI > 3,000 IA). Importantly, none of the aneurysms in our study cohort were included in the training datasets of either platform, ensuring true external validation. Pre-trained 12-layer convolutional neural network models performed: (1) intracranial vessel segmentation; (2) aneurysm detection; (3) neck width/length calculation. No manual adjustments were allowed.

Statistical analysis

Continuous variables were reported as mean ± SD if normally distributed, or as median (IQR) otherwise, based on the Shapiro-Wilk test. Categorical variables were described as n (%). Reproducibility was evaluated via CV. Bland-Altman analysis assessed consistency between methods (reported as mean difference [MD] and 95% limits of agreement [LOA]); 95% LOA within ± 1.0 mm was clinically acceptable [5].Analyses were performed using Statistical Package for the Social Sciences (SPSS) 22.0 (IBM) and R 4.2.0 (blandr package). Two-tailed P < 0.05 was statistically significant.

Results

Baseline characteristics of the study population

A total of 148 patients (harboring 163 IA) were included in the CTA-based measurement analysis; of these, 86 patients formed the subgroup for comparative analysis against DSA.Baseline demographics and IA anatomical distributions are summarized. (Table 1).

Table 1.

Baseline characteristics of the study population

Characteristic	Number (n)	Proportion (%)
Gender
Female	106	71.6
Male	42	28.4
Age group (years)
≤ 40	25	16.9
41–60	90	60.8
> 60	33	22.3
Aneurysm location
MCA	33	20.2
AComA	33	20.2
ICA-PComA	65	39.9
Basilar Artery Bifurcation	13	8.0
Other locations	19	11.7

The cohort was predominantly female (71.6%). The largest age group was 41–60 years (60.8%). The most common site was the Internal Carotid Artery-Posterior Communicating Artery (ICA-PComA) (39.9%), followed by the Middle Cerebral Artery (MCA) and Anterior Communicating Artery (AComA) (Fig. 2).

Fig. 2.

Comparative evaluation of AI-based platforms against digital subtraction angiography for visualizing an anterior communicating artery aneurysm. Representative images from the (a, b) Shukun AI platform, (c, d) United Imaging AI platform, and (e, f) reference standard DSA are shown

Descriptive statistics of aneurysm measurements across different methods

Descriptive statistics for aneurysm neck width and maximum length measurements are presented in Table 2.

Table 2.

Descriptive statistics of aneurysm measurements by different methods

Measurement Method (Measured Parameter)	Range (mm)	Mean ± SD (mm)	CV
A1Shukun AI (Neck Width)	2.33–7.30	4.51 ± 1.40	0.31
UIH AI (Neck Width)	2.27–9.55	4.53 ± 1.62	0.36
Manual Measurement 1 (Neck Width)	1.13–9.70	5.18 ± 1.87	0.36
Manual Measurement 2 (Neck Width)	0.15–10.92	5.23 ± 2.21	0.42
DSA (Neck Width)	2.63–6.37	4.15 ± 0.90	0.22
Shukun AI (Maximum Length)	5.70–18.60	11.12 ± 2.60	0.23
UIH AI (Maximum Length)	4.14–18.80	12.06 ± 3.14	0.26
Manual Measurement 1 (Maximum Length)	4.71–18.95	11.57 ± 2.67	0.23
Manual Measurement 2 (Maximum Length)	5.32–19.21	11.65 ± 2.75	0.24
DSA (Maximum Length)	6.78–14.22	10.66 ± 1.54	0.14

As shown in Fig. 3, the CV differed across measurement methods. Shukun AI achieved the highest measurement consistency among CTA-based methods, whereas DSA demonstrated the lowest overall variability, confirming its superior reproducibility.

Fig. 3.

Comparison of coefficient of variation among different measurement methods for intracranial aneurysm parameters

Bland-Altman analysis

Bland–Altman analysis revealed variable agreement among the measurement methods (Figs. 4, 5 and 6).

Fig. 4.

Bland-Altman analysis of inter-AI platform agreement for aneurysm neck width and maximum length measurements

Fig. 5.

Agreement between AI platforms and DSA reference standard for aneurysm neck width and maximum length measurements

Fig. 6.

Intra-observer consistency of manual measurements and agreement with DSA for aneurysm neck width and maximum length

Between the two AI platforms, Shukun AI and UIH AI showed close agreement for both neck width and maximum length, with minimal mean bias. However, when compared with the DSA reference, both AI systems and manual measurements exhibited wider limits of agreement, exceeding the predefined ± 1.0 mm clinical threshold.Among all methods, manual intra-observer consistency was the highest.

Discussion

Summary of key findings

This study systematically evaluated the accuracy and consistency of two commercial AI platforms (Shukun AI and UIH AI) and manual measurements in quantifying intracranial aneurysm morphological parameters, using DSA as the reference standard. Three principal findings emerged: (1) Among CTA-based measurements of 163 intracranial aneurysms, Shukun AI demonstrated the lowest coefficient of variation for both neck width (0.31) and maximum length (0.23), outperforming UIH AI (CVs: 0.36 and 0.26) and manual measurements (CVs: 0.36–0.42 and 0.23–0.24); (2) DSA exhibited the lowest variability across all methods (CVs: 0.22 for neck width, 0.14 for maximum length), confirming its status as the gold standard; (3) Critically, in the subgroup of 86 patients who underwent both CTA and DSA within one week, the 95% LOA for all measurement pairs—including inter-AI comparisons, AI versus DSA, and manual versus DSA—fell outside the clinically acceptable threshold of ± 1.0 mm, indicating that current measurement approaches have inherent precision limitations that may impact clinical decision-making.

AI’s advantage in measurement reproducibility

Deep learning algorithms have demonstrated robust performance in intracranial aneurysm detection, a capability that has garnered broad recognition [12–14]. Our findings extend this evidence by confirming that AI platforms exhibit superior reproducibility compared to manual measurements. Shukun AI’s CVs (0.31 for neck width, 0.23 for maximum length) were significantly lower than those of manual measurements (0.36–0.42 and 0.23–0.24, respectively), consistent with prior investigations. Planinc et al. [9] reported that all human raters, including senior radiologists, demonstrated standard deviations exceeding the 1 mm threshold for aneurysm growth, whereas AI-based approaches achieved performance comparable to senior raters in interval aneurysm monitoring. Similarly, a phantom-based validation study [15] found that computer-assisted semi-automated measurement yielded errors of only 1.1–2.5%, compared to 1.7–19.1% for manual measurement, with a concordance correlation coefficient of 0.993.

These findings collectively suggest that AI can effectively reduce operator-dependent measurement variability. However—and this is the critical issue our study reveals—improved reproducibility does not automatically equate to improved accuracy when validated against the gold standard.

Systematic biases in AI versus DSA measurements

The two commercial AI systems exhibited distinct systematic biases in aneurysm measurement.The observed pattern—UIH AI tending to overestimate and Shukun AI to underestimate—likely reflects fundamental differences in algorithmic design and training data.UIH AI appears to apply a broader HU threshold during lumen segmentation, incorporating partial-volume voxels near vessel walls and thereby generating slightly larger diameters [16].In contrast, Shukun AI adopts a more conservative surface extraction strategy with higher HU thresholds and stricter boundary clipping, resulting in smaller measurements.

Additionally, variations in training datasets—including scanner models, imaging protocols, and annotation standards—may further influence model calibration and bias direction [17].

Recognizing these algorithmic and data-related factors is essential for achieving cross-platform harmonization and establishing clinically consistent AI measurement standards [18].

Previous research has shown that computer-assisted methods often yield larger neck diameter estimates than manual measurements, likely due to more accurate detection of the maximal vessel boundary [19], which may partially explain UIH AI’s overestimation trend.

Conversely, a monocentric comparison between CTA and DSA [20] reported that, although mean aneurysm and parent vessel diameters were similar overall, DSA tended to yield smaller measurements for microaneurysms (< 3 mm).These findings highlight that measurement discrepancies arise not only from algorithmic differences but also from intrinsic variations between imaging modalities—including spatial resolution, contrast dynamics, and temporal acquisition characteristics—underscoring the complexity of achieving cross-modality precision in aneurysm quantification.

The challenge of the ± 1 mm clinical acceptability threshold

A pivotal finding of our study is that the 95% LOA for all measurement pairs failed to fall within the pre-specified clinically acceptable range of ± 1.0 mm. This has profound implications for clinical translation, as aneurysm growth ≥ 1 mm is widely regarded as a surrogate marker for elevated rupture risk and serves as a critical determinant of treatment strategy and long-term surveillance protocols [21–23].

Planinc et al. [9]similarly demonstrated that all raters—including senior radiologists—exhibited standard deviations exceeding the 1 mm threshold, indicating that achieving this level of precision is not merely an AI optimization problem but reflects inherent limitations of current imaging modalities. A recent scoping review [24] emphasized that existing deep learning studies remain at the stage of technical validation, with high risks of bias and limited clinical applicability due to retrospective designs, convenience sampling, and exclusion of ruptured or previously treated aneurysms. Our measurement consistency analysis further corroborates this assessment, revealing that the field has not yet progressed from technical validation to real-world clinical performance evaluation.

Moreover, a volumetric AI measurement study [2] demonstrated that AI-based volumetric assessment detected growth in all aneurysms that ruptured during conservative management, including cases missed by manual linear measurements. This suggests that traditional linear parameters may be inherently insensitive for detecting subtle but clinically significant morphological changes, warranting future investigation of volumetric and hemodynamic parameters.

Operator-dependent variability in manual measurements

Manual measurements exhibited higher variability in our study. The CVs for neck width were 0.36 (measurement 1) and 0.42 (measurement 2), reflecting intra-rater variability even with a one-month interval. Kim et al. [25] found that neck width measurements demonstrated higher intra- and inter-observer variability compared to height and width measurements in both CTA and magnetic resonance angiography, corroborating our finding that manual measurement 1 versus DSA yielded the poorest agreement for neck width (MD: -1.62 mm, 95% LOA: -4.87 to 1.62 mm).

The difficulty in measuring neck width likely stems from: (1) subjective definition of the neck plane; (2) geometric radiographic distortion; and (3) anatomical complexity at the aneurysm-vessel junction. Munarriz et al. [19] reported that the VMTKlab software achieved only moderate agreement with DSA for neck width (ICC 0.57), whereas the other three morphological parameters showed good agreement (ICC 0.85–0.89), further confirming the universal challenge of neck plane determination.

Importantly, our data revealed that intra-rater variability in manual measurements (MD between measurement 1 and 2: 0.05 mm for neck width with 95% LOA: -2.20 to 2.30 mm) exceeded the 1 mm threshold in many instances, indicating that even the same experienced radiologist cannot consistently reproduce measurements within clinically acceptable limits when assessing the same aneurysm at different time points. This temporal inconsistency poses a significant challenge for serial surveillance, where distinguishing true aneurysm growth from measurement error is paramount.

Imaging modality contributions

Sui et al. [20] found that DSA could clearly visualize the relationship between the aneurysm neck and parent artery in all cases, whereas CTA had a 90% accuracy rate. This indicates that DSA’s superior spatial resolution and hemodynamic information provide an advantage in delineating neck boundaries. The lowest CVs of DSA in our study (0.22 for neck width, 0.14 for maximum length) partially reflect this characteristic.

However, CTA has become the preferred modality for initial screening and serial monitoring due to its noninvasive nature, whereas DSA is associated with risks including contrast-induced nephropathy and vascular complications [12, 26]. This underscores the urgency of improving CTA measurement accuracy, including AI-assisted approaches, to enable safer longitudinal surveillance.

Clinical implications for treatment decision-making

The clinical significance of neck width measurement discrepancies cannot be overstated. Wide-necked aneurysms (neck width ≥ 4 mm) typically require adjunctive devices (stents or balloons) to ensure complete embolization and prevent coil migration into the parent artery [27]. The systematic biases observed in our study—with maximum discrepancies reaching ± 1.6 mm—could affect threshold determination. For instance, if an aneurysm with an actual neck width of 3.5 mm is overestimated by UIH AI to 4.3 mm, unnecessary stent placement may be undertaken; conversely, if underestimated by Shukun AI to 2.5 mm, the risk of incomplete occlusion increases.

Furthermore, aneurysm growth ≥ 1 mm is a pivotal criterion for rupture risk stratification. A retrospective analysis of five aneurysms that ruptured during conservative management [2] found that AI-based volumetric measurements were more sensitive than manual linear measurements, detecting growth in all ruptured cases, including those undetected by manual assessment. This suggests that volumetric parameters may be superior to linear metrics, yet current commercial AI platforms primarily focus on linear measurements, with volumetric assessment remaining inadequately validated.

The heterogeneity in aneurysm characteristics further complicates measurement interpretation. Korja et al. [21] analyzed a consecutive series of 1,993 ruptured intracranial aneurysms and found that 68% had maximum diameters < 10 mm, with median size varying considerably by location (e.g., ophthalmic artery: 11 mm vs. pericallosal artery: 6 mm). This anatomical diversity, combined with our finding that neither AI nor manual measurements achieved ± 1 mm agreement with DSA, underscores the need for location-specific validation of AI algorithms and potentially adaptive measurement protocols based on aneurysm characteristics.

Critical implications for serial aneurysm surveillance

A particularly important practical implication of our findings concerns the longitudinal monitoring of unruptured intracranial aneurysms. Our Bland-Altman analysis demonstrated that inter-platform measurement variability between Shukun AI and UIH AI yielded 95% LOA of -3.12 to 3.09 mm for neck width and − 3.37 to 5.24 mm for maximum length—ranges that substantially exceed the ± 1 mm threshold used to define aneurysm growth. Similarly, the manual intra-rater variability (95% LOA: -2.20 to 2.30 mm for neck width) indicates that even when the same experienced radiologist re-evaluates an aneurysm at different time points, measurement discrepancies can exceed 1 mm in a considerable proportion of cases.

These findings have critical implications for clinical practice: switching between different AI platforms during serial surveillance—or even the same operator at different time points—may introduce measurement variability that mimics or obscures true aneurysm growth. For example, an aneurysm measured as 5.0 mm by Shukun AI at baseline might be measured as 6.5 mm by UIH AI at one-year follow-up, falsely suggesting 1.5 mm growth (and thus elevated rupture risk) when the discrepancy may simply reflect inter-platform bias. Conversely, true aneurysm growth could be masked if measurement variability coincidentally compensates for actual size increase.

This phenomenon has been underrecognized in clinical practice and warrants explicit acknowledgment. To minimize spurious growth assessments and ensure reliable detection of true morphological change, we strongly recommend that serial imaging surveillance of the same aneurysm should consistently employ the same AI platform, ideally using identical imaging protocols and reconstruction parameters. Institutions implementing AI-assisted aneurysm surveillance should establish standardized workflows that ensure measurement consistency across time points, and radiologists should be aware that inter-platform or intra-rater measurement variability can exceed clinically meaningful thresholds. Future guidelines for aneurysm surveillance protocols should incorporate these considerations to prevent misinterpretation of interval changes.

Limitations and future directions

This study has several limitations. First, it was conducted at a single center, which may introduce equipment- and operator-related bias and limit the generalizability of the findings. Second, only two commercial AI platforms were evaluated; thus, the conclusions may not extend to other commercial or open-source systems.Third, the cross-sectional design precluded longitudinal assessment of aneurysm growth dynamics and patient outcomes.

Future studies should address these limitations by including multicenter cohorts with diverse scanners, imaging protocols, and AI vendors to improve external validity. Longitudinal research is warranted to determine whether inter-method measurement discrepancies influence treatment planning or surveillance strategies. Moreover, future developments in aneurysm quantification are expected to transition from linear to volumetric assessment, as volumetric AI approaches can capture subtle three-dimensional morphological changes that precede detectable diameter growth. Integrating automated 3D segmentation, standardized reconstruction protocols, and transparent model validation will be essential to establish reliable, reproducible, and clinically meaningful AI-based aneurysm monitoring frameworks.

Conclusion

This study demonstrates that commercial AI platforms achieve superior measurement reproducibility compared to manual assessment. However, when validated against DSA, both AI platforms and manual measurements failed to meet the clinically critical ± 1.0 mm agreement threshold. Critically, inter-platform measurement variability substantially exceeds the ± 1 mm threshold for defining aneurysm growth, necessitating consistent measurement methodology across serial surveillance studies to avoid spurious growth signals or missed true progression.

Abbreviations

CTA

Computed Tomography Angiography

DSA

Digital Subtraction Angiography

UIH

United Imaging Healthcare

IA

Intracranial Aneurysm

AI

Artificial

CV

Coefficient of Variation

LOA

Limits of Agreement

MD

Mean

IRB

Institutional Review Board

AVMs

Arteriovenousmalformations

HU

Hounsfield unit

ICA-PComA

Internal Carotid Artery-Posterior Communicating Artery

MCA

Middle Cerebral Artery

AComA

Anterior Communicating Artery

Author contributions

Conceptualization: C.Y.; Methodology: C.Y.,X.D.; Validation: C.Y.,X.D.;Formal analysis: S.D., L.H.; Investigation: S.D., L.H; Writing - Original Draft: C.Y.; Writing - Review & Editing: X.D; All authors have read and agreed to the published version of the manuscript.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This retrospective study was approved by the Institutional Review Board of Zhuzhou Central Hospital (approval number: ZZCH-IRB-2022-073; approval date: August 15, 2022). The requirement for informed consent was waived by the ethics committee due to the retrospective nature of the study and the use of fully anonymized data. All procedures were performed in accordance with the Declaration of Helsinki (2013 revision) and relevant institutional guidelines for medical research involving human subjects.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.