Multicenter-derived machine learning model for individualized rupture-risk prediction in intracranial aneurysms: study protocol and preliminary validation

Introduction

Epidemiological studies indicate a global intracranial aneurysms (IAs) prevalence of approximately 3.2% among individuals aged 50 years. Despite an annual rupture rate of merely 0.25–2.00%, IAs rupture-induced subarachnoid hemorrhage (SAH) carries a mortality rate of 25–50%, with approximately half of survivors sustaining permanent neurological impairment (1). This triad of high prevalence, low rupture rate, yet high disability and mortality rates establishes IAs as a critical unresolved clinical challenge in neurological disorders. Recently, the detection rates of asymptomatic unruptured intracranial aneurysms (UIAs) have substantially increased due to advancements in medical imaging technology and population aging. Current guidelines highlight persistent knowledge gaps regarding optimal treatment selection for IAs patients, leading to dual challenges: unnecessary overtreatment in low-risk cases versus missed therapeutic windows in high-risk individuals (2). Therefore, establishing an accurate rupture risk prediction system is essential for optimizing IAs management. The rupture risk of IAs is influenced by the complex interplay among various factors, including clinical, morphological, hemodynamic, and radiomics factors, demonstrating significant individual heterogeneity. Therefore, single-factor assessment is insufficient to guide treatment strategies.

Logistic regression (LR) remains one of the most widely used and conventional methods for developing predictive models, owing to its simplicity and intuitive interpretability. In recent years, however, advances in artificial intelligence (AI) have enabled researchers to construct various models for predicting IAs rupture risk using diverse types of patient data, with remarkable results achieved (3,4). Machine learning (ML), which automates model building through data-driven learning, can significantly improve the accuracy (ACC) and efficiency of decision-making. ML comprises a wide range of algorithms, each with distinct advantages. It should be noted that the ACC of intelligent models is highly dependent on large datasets and the natural history characteristics of IAs. Before these models can be translated into clinical practice, personalized management strategies for IAs patients and further optimization of intelligent predictive tools must be substantiated through extensive clinical validation.

Therefore, in this study, we integrated clinical, computed tomographic angiography (CTA) morphological, and radiomic features of patients with IAs to construct a predictive model for rupture risk using LR. Subsequently, a visual nomogram was constructed to quantify the impact of each variable on the rupture risk. Finally, to ensure methodological robustness and minimize algorithmic bias, we conducted rigorous comparative analyses of seven ML approaches using the identical feature set, thereby enabling objective evaluation of model performance across different computational paradigms. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2593/rc).

Methods

Patient selection

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The Second Affiliated Hospital of Soochow University (No. JD-LC2024029-I01), and the requirement for informed consent was waived because this was a retrospective study of preexisting imaging data. The other participating centers were informed of and agreed to this study. Consecutive patients clinically suspected of cerebrovascular diseases who underwent cranial CTA at The Second Affiliated Hospital of Soochow University (Center I) between January 2020 and November 2024 were retrospectively enrolled. Among 404 patients (450 aneurysms, n=450) meeting inclusion and exclusion criteria with confirmed IAs, random allocation at a 7:3 ratio yielded training set (n=314) and internal validation set (n=136). Simultaneously, eligible IAs patients from The First Affiliated Hospital of Soochow University (Center II) and Zhongda Hospital Southeast University (Center III) during the same period constituted external validation set I (125 patients; 148 aneurysms, n=148) and external validation set II (227 patients; 279 aneurysms, n=279), respectively. The inclusion criteria for patients were as follows: (I) patients undergoing cranial CTA with confirmed IAs; (II) image quality meeting postprocessing and feature extraction requirements. The exclusion criteria were as follows: (I) traumatic or secondary IAs; (II) prior history of IAs intervention; (III) fusiform aneurysms, dissecting aneurysms, or vascular malformations; and (IV) incomplete clinical or imaging data. Figure 1 shows the study flowchart.

Figure 1 Flowchart of patient selection and dataset division for the development and validation of the IAs rupture prediction model. CTA, computed tomographic angiography; IAs, intracranial aneurysms; LR, logistic regression.

Imaging and clinical data

A multicenter CTA protocol was implemented. Center I utilized three scanners [Philips IQon Spectral CT and iCT (Philips, Amsterdam, Netherlands); GE Revolution CT (GE Healthcare, Chicago, IL, USA)], whereas Centers II and III used one scanner each (Philips IQon Spectral CT and GE Discovery CT750 HD, respectively). All acquisitions ensured full brain coverage (skull base to vertex). Comprehensive technical details are listed in Table S1.

Clinical data were extracted from electronic medical records, including the following: age (years), sex, history of hypertension, admission systolic blood pressure (mmHg), history of diabetes, admission blood glucose (mmol/L), history of hyperlipidemia, total cholesterol (mmol/L), and triglycerides (mmol/L).

The rupture status of IAs was determined according to the following criteria: (I) for patients with SAH, an aneurysm was classified as ruptured if it was the only aneurysm adjacent to the blood clot. Conversely, when non-adjacent aneurysms were present or multiple aneurysms (≥2) coexisted, surgical confirmation was required to identify the ruptured lesion; and (II) asymptomatic patients without SAH were classified as having unruptured aneurysms (5).

Morphological features measurement

Cranial CTA images were exported from the Picture Archiving and Communications System (PACS) in Digital Imaging and Communication in Medicine (DICOM) format and uploaded to AneurDoc (version 2.0; Shukun Technology Co., Ltd., Beijing, China). Morphological parameters for each IA were automatically computed using the “Advanced Aneurysm” software function. The analyzed parameters included (1,6,7): multiplicity, location, height, diameter, width, maximum diameter, neck diameter, neck area, parent artery average diameter, size ratio (SR), aspect ratio (AR), height:width ratio, volume:neck area ratio, aneurysm angle (AA), flow angle (FA), undulation index (UI), non-sphericity index (NSI), and ellipticity index (EI). Precise definitions and formulae for these parameters are detailed in Table S2.

Radiomics feature selection

Three-dimensional (3D) aneurysm segmentation on cranial CTA images was performed using AneurDoc (version 2.0; Shukun Technology Co., Ltd.), a knowledge-augmented convolutional neural network (CNN)-based system comprising two cascaded networks: (I) msResU-Net, a 3D ResU-Net with multiscale feature extraction for simultaneous vessel and aneurysm segmentation, and (II) mfResNet, a 3D ResNet-18-based diagnostic model that incorporates prior anatomical knowledge to reduce false positives. All segmented aneurysm regions were reviewed and manually adjusted by two senior radiologists to ensure ACC, with the refined contours defined as regions of interest (ROIs) for subsequent analysis (8). The CTA images and ROIs were then transferred to a cloud-based radiomics platform (Shukun Technology). To address multi-scanner variability, all images underwent preprocessing including resampling and gray-level discretization before feature extraction. A total of 1,874 radiomics features were extracted from each ROI, encompassing first-order statistics, gray-level co-occurrence matrix (GLCM), gray level dependence matrix (GLDM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), neighboring gray tone difference matrix (NGTDM), and shape-based features. All features were normalized using z-score standardization for inter-scanner consistency (9). Feature selection was conducted within the training set. Duplicate and zero-variance features were first excluded. Redundant features were then eliminated through Pearson correlation analysis (absolute correlation coefficient ≥0.9). Subsequently, the SelectKBest method was applied based on univariate analysis (P<0.05) to retain statistically relevant features. Finally, the least absolute shrinkage and selection operator (LASSO) algorithm was employed for further dimensionality reduction to identify the most predictive radiomics features for model construction. The radiomics score (Radscore) was computed as: Radscore = (∑βⱼ × Xⱼ) + intercept, where Xⱼ denotes the value of the j-th selected feature and βⱼ represents its corresponding coefficient (10). The overall workflow is shown in Figure 2.

Figure 2 Workflow depicting the key steps from image segmentation, feature selection, model construction, and validation. AUC, area under the curve; C, clinical; DCA, decision curve analysis; DT, decision tree; GLCM, gray-level co-occurrence matrix; GLDM, gray level dependence matrix; GLRLM, gray-level run length matrix; GLSZM, gray-level size zone matrix; LR, logistic regression; M, morphological; NB, naive bayes; NGTDM, neighboring gray tone difference matrix; PA, passive-aggressive; R, radiomics; ROI, region of interest; SGD, stochastic gradient descent; SVC, support vector classifier; SVM, support vector machine.

Prediction models construction and evaluation

In the training set, univariate analysis of clinical and morphological parameters identified variables with significant differences (P<0.05), which were subsequently included in multivariate LR to determine independent risk factors for IAs rupture (P<0.05). These predictors, together with selected radiomics features, were incorporated into the cloud-based radiomics platform (Shukun Technology) for modeling. Leveraging its clinical prevalence and interpretability (11), LR was used to develop six distinct models via 10-fold cross-validation: Model_M (morphological), Model_R (radiomics), Model_C&M (clinical-morphological), Model_C&R (clinical-radiomics), Model_M&R (morphological-radiomics), and Model_C&M&R (integrated clinical-morphological-radiomics). Each model was evaluated in the training and internal validation sets, with the best-performing model further validated in two external cohorts (external validation sets I and II). Performance was assessed using receiver operating characteristic (ROC) analysis, with the area under the curve (AUC) as the primary metric, supplemented by sensitivity (SEN), specificity (SPE), and ACC. Model comparisons employed DeLong’s test for AUC differences. Calibration curves evaluated prediction ACC, and decision curve analysis (DCA) quantified clinical utility across decision thresholds. A nomogram integrating the Radscore with independent clinical and morphological risk factors was developed to visualize rupture risk.

Multiple ML models construction and validation

ML comprises a diverse array of algorithms, each offering distinct advantages. To thoroughly assess the stability of predictive performance across multiple ML approaches, this study incorporated seven classical ML algorithms—support vector machine (SVM), adaptive boosting (AdaBoost), stochastic gradient descent (SGD), linear support vector classifier (SVC), Gaussian naïve bayes (NB), passive-aggressive (PA), and decision tree (DT)—based on the feature set of Model_C&M&R. Following parameter optimization for each algorithm within the training set, their stability was evaluated on the internal validation set to identify the best-performing model. The selected model was further validated for its cross-center generalizability using external validation sets I and II.

Statistical methods

Statistical analyses of clinical and morphological variables were performed using SPSS software (version 26.0; IBM Corp., Armonk, NY, USA). The normality of continuous variables was assessed using the Shapiro-Wilk test. Normally distributed continuous variables were expressed as mean ± standard deviation (mean ± SD) and compared between groups using the independent samples t-test. Non-normally distributed variables were summarized as median [interquartile range (IQR)] and compared using the Mann-Whitney U test. Categorical variables were presented as numbers (percentages) and analyzed using the Chi-squared test or Fisher’s exact test, as appropriate. A two-sided P<0.05 was considered statistically significant. Variables exhibiting significant associations (P<0.05) in univariate analysis were subsequently entered into multivariate LR to identify independent risk factors, with results reported as odds ratio (OR) and corresponding confidence intervals (CIs).

Results

Clinical factors and morphological parameters

This multicenter retrospective study included a total of 756 patients with IAs, comprising 877 aneurysms. The cohort was allocated into a training set (n=314), an internal validation set (n=136), and two external validation sets (external validation set I, n=148; external validation set II, n=279), the baseline characteristics and outcome distributions of the training, internal validation, and external validation cohorts are summarized in Table S3. Univariate analysis of clinical and morphological parameters within the training set (Table 1) revealed several significant differences between ruptured and unruptured groups. Specifically, the ruptured group exhibited a lower proportion of female patients and a higher prevalence of diabetes mellitus. Additionally, patients in the ruptured group were significantly younger and had higher admission blood glucose levels. Several morphological parameters were also significantly associated with rupture risk, including aneurysm location, diameter, width, height, maximum diameter, FA, parent artery average diameter, IU, EI, SR, AR, and volume-to-neck area ratio (all P<0.05). No statistically significant differences were observed in the remaining clinical or morphological variables between the two groups (P>0.05).

Multivariate analysis (Table 2) identified the following factors as independent predictors of IAs rupture: age (OR =0.948; 95% CI: 0.927–0.971), admission blood glucose (OR =1.299; 95% CI: 1.171–1.440), aneurysm location in the anterior circulation (excluding ICA) (OR =3.130; 95% CI: 1.450–6.753), aneurysm diameter (OR =1.433; 95% CI: 1.238–1.658), and parent artery average diameter (OR =0.674; 95% CI: 0.470–0.955) (all P<0.05). These variables were subsequently incorporated into the development of predictive models.

Radiomics features selection

A total of 1,874 radiomics features were initially extracted from each aneurysm. Following feature preprocessing, duplicate features (n=253) and features with zero variance (n=1) were first excluded. Subsequently, features exhibiting high collinearity (absolute Pearson correlation coefficient ≥0.9; n=943) were also removed. Finally, 10 radiomics features significantly associated with IAs rupture risk were selected using the SelectKBest univariate analysis followed by LASSO regression, and were employed for subsequent model construction. The Radscore was computed as a linear combination of the selected features weighted by their respective coefficients derived from the LASSO model. Detailed feature descriptions and corresponding coefficients are summarized in Table 3 and were incorporated into the development of subsequent prediction models.

Prediction model establishment and validation

Based on multivariate analysis, all model metrics are summarized in Table 4, with corresponding ROC curves presented in Figure 3A,3B. Among the models evaluated, Model_C&M&R demonstrated the best overall performance. DeLong’s test showed no significant AUC difference between Model_C&M&R and Model_C&M (P>0.05), yet the net reclassification improvement (NRI) for Model_C&M&R was 72.49% (training) and 36.74% (validation), indicating significantly improved reclassification. DCA confirmed higher net clinical benefit for Model_C&M&R within a threshold probability range of 4–85% (Figure 3C). Calibration was satisfactory (Brier score =0.090; Figure 3D). In external validation, Model_C&M&R achieved AUC =0.773 (95% CI: 0.687–0.859) in set I and AUC =0.735 (95% CI: 0.676–0.794) in set II. The corresponding confusion matrix is shown in Figure 4. A nomogram incorporating six predictors—age, admission glucose, diameter, location, parent artery diameter, and Radscore—was developed for clinical use (Figure 5).

Figure 3 ROC curves of six models in training set (A) and internal valuation set (B). DCA curve of Model_C&M&R in internal validation set (C). Calibration curve of Model_C&M&R in internal validation set (D). AUC, area under the curve; C, clinical; DCA, decision curve analysis; M, morphological; R, radiomics; ROC, receiver operating characteristic.

Figure 4 Confusion matrix used for the internal validation set (A), external validation set I (B) and external validation set II (C).

Figure 5 A nomogram for individualized prediction of IAs rupture risk, incorporating six predictors: age, admission blood glucose, location, aneurysm diameter, parent artery average diameter, and Radscore. Location: 0, ICA; 1, MCA; 2, anterior circulation; 3, posterior circulation. IAs, intracranial aneurysms; Radscore, radiomics score.

Construction and validation of multiple ML risk assessment models

Building upon Model_C&M&R, seven additional classical ML algorithms—SVM, AdaBoost, SGD, linear SVC, Gaussian NB, PA, and DT—were employed to develop predictive models. The performance metrics of these models within the training and internal validation sets are summarized in Table 5 and visually presented in Figure 6. With the exception of Gaussian NB (AUC =0.748; 95% CI: 0.635–0.861) and DT (AUC =0.686; 95% CI: 0.554–0.817), all other models demonstrated strong predictive performance based on AUC values: SVM (AUC =0.885; 95% CI: 0.817–0.953), SGD (AUC =0.910; 95% CI: 0.848–0.972), AdaBoost (AUC =0.898; 95% CI: 0.831–0.965), linear SVC (AUC =0.900; 95% CI: 0.839–0.961), and PA (AUC =0.903; 95% CI: 0.840–0.967). Among these, the SGD model performed optimally across multiple evaluation metrics, matching the LR model in AUC (both 0.910). According to DeLong’s test, no statistically significant difference was observed between the AUC values of the LR and SGD models (P>0.999). DCA revealed that both models provided substantial net clinical benefits across a wide range of threshold probabilities, supporting their potential utility in clinical practice. Notably, within the threshold probability interval of approximately 4–84%, the LR model yielded higher net benefit compared to the SGD model (Figure 6C). Furthermore, calibration curve analysis indicated that the SGD model (Brier score =0.098) exhibited greater deviation between predicted and observed probabilities than the LR model (Brier score =0.090), suggesting inferior calibration performance for the SGD model (Figure 6D). In conclusion, although the LR and SGD models did not differ significantly in discriminative ability (as measured by AUC), the LR model demonstrated superior calibration stability and greater net clinical benefit, supporting its preference for clinical application based on comprehensive evaluation.

Figure 6 ROC curves of different ML models in training set (A) and internal validation set (B). DCA curve (C) and calibration curve (D) of LR and SGD models in internal validation set. AdaBoost, adaptive boosting; AUC, area under the curve; DCA, decision curve analysis; DT, decision tree; LR, logistic regression; ML, machine learning; NB, naive bayes; PA, passive-aggressive; ROC, receiver operating characteristic; SGD, stochastic gradient descent; SVC, support vector classifier; SVM, support vector machine.

Discussion

In this study, age, admission blood glucose, aneurysm diameter, location, parent artery average diameter, and 10 radiomics features were identified as independent predictors of IAs rupture.

Although numerous studies have established advanced age as an independent risk factor for IAs rupture—as reported by Wermer et al. (12) and Shi et al. (13)—our findings revealed a contrasting association. The median age of the ruptured group was significantly lower than that of the unruptured group (58.0 vs. 68.0 years), with each one-year increase in age corresponding to a 4.8% decrease in rupture risk. We hypothesize that this inverse relationship may be explained by more active collagen metabolism in younger patients, as well as a tendency for unruptured IAs to stabilize over time—consistent with Nahed et al. (14). Univariate analysis indicated a significantly higher prevalence of diabetes history in the ruptured group (58.2% vs. 25.5%). However, multivariate analysis did not identify diabetes history as an independent predictor, suggesting that its effect may be mediated through direct metabolic measures such as blood glucose levels. Admission blood glucose levels were significantly elevated in the ruptured group (8.02 vs. 5.39 mmol/L) and were confirmed as an independent risk factor. It remains unclear whether hyperglycemia is a cause or a consequence of rupture, as SAH can itself induce physiologic stress and elevate glucose. Evidence from previous studies indicates a significant correlation between IAs size and rupture risk. van der Kamp et al. (15) demonstrated that a diameter ≥7 mm serves as an independent predictor. Our study supports these findings, confirming aneurysm diameter as an independent risk factor, with the median diameter in the ruptured group being significantly larger (4.65 vs. 3.36 mm). Aneurysm location was also an independent risk factor (16-21). Although ICA was used as the reference group in our multivariate analysis, this does not imply a low rupture risk for ICA aneurysms, which may be explained by hemodynamic factors: the curved architecture of the ICA can accelerate blood flow velocity and promote turbulent flow within the aneurysm (22). In our results, anterior circulation location (excluding ICA) showed an even higher risk compared to ICA (OR =3.130; P=0.004), further supporting the clinical importance of location as a risk factor. Furthermore, a smaller parent artery diameter was associated with increased rupture risk. Previous studies have suggested that narrower parent arteries may result in elevated blood flow velocity and wall shear stress, thereby augmenting mechanical stress on the aneurysm wall (23-25).

Multiple studies have demonstrated significant differences in radiomics features between ruptured and unruptured IAs (26-29). In our study, 10 optimal features were selected from 1,874 imaging-derived features. Shape_Sphericity was identified as the radiomics feature most strongly associated with IAs rupture, with a coefficient of −0.607, indicating that decreased sphericity is associated with higher risk. This result is consistent with Liu et al. (27) and aligns with the recognized importance of morphological irregularity in predicting aneurysm rupture (30,31).

As LR remains one of the most widely used methods for developing prediction models (32), this study employed the LR algorithm to construct six distinct models. Model_C&M&R, integrating clinical, morphological, and radiomics features, achieved the highest discriminative performance with AUCs of 0.887 in the training set and 0.910 in the internal validation set. Although DeLong’s test indicated no statistically significant improvement in AUC from adding radiomics features, the significant cumulative NRI demonstrated that Model_C&M&R provided a positive predictive refinement over Model_C&M.

A nomogram is a graphical tool that provides an intuitive representation of complex statistical models, enabling straightforward calculation of the probability of a specific event. It has been widely adopted in both clinical practice and research. Based on the Model_C&M&R, we further developed a nomogram incorporating six independent predictors. The prediction model derived from this nomogram achieved an AUC of 0.879 in the internal validation set and demonstrated excellent performance in two independent external validation sets. This model quantifies the contribution of each risk factor to the probability of IAs rupture and offers a more direct and intuitive assessment of rupture risk, thereby simplifying and facilitating the evaluation of IAs rupture risk in clinical settings.

To evaluate stability across ML methods, seven additional classical ML algorithms were tested. All models except Gaussian NB and DT exhibited robust performance. The SGD model performed comparably to LR in discrimination (AUC =0.910), but was inferior in calibration and clinical net benefit across most threshold probabilities. Thus, the LR model was considered optimal. Consensus on the optimal algorithm remains elusive, as evidenced by varying top-performing algorithms in other studies (3,4), indicating the need for further exploration.

Limitations

This study has several limitations. First, the low incidence of IAs rupture led to a relatively limited number of ruptured cases and an imbalanced sample size between groups. Although consistent with epidemiological characteristics, this imbalance may increase overfitting risk. We employed 10-fold cross-validation to mitigate this and enhance model generalizability. Second, although previous studies suggest morphological changes may occur after IAs rupture (33-35), ethical constraints made it unfeasible to obtain pre- and post-rupture morphological data. This factor could not be incorporated and will be addressed in future prospective studies. Third, as a multi-center study, images in the external validation set were acquired using different scanners and protocols, potentially introducing variability. To minimize inter-site heterogeneity, all images underwent resampling and gray-level discretization during preprocessing. Finally, further research is warranted to explore the clinical translational potential and practical utility of the proposed model, aiming to provide a more reliable tool for clinical decision-making.

Conclusions

The integrated model combining clinical, CTA morphological, and radiomics features significantly enhances the predictive ACC for IAs rupture risk. The subsequently developed nomogram and its associated predictive model allow for simple and intuitive assessment of IAs rupture risk, while also facilitating personalized risk stratification for individual patients. Although various ML algorithms demonstrated competent predictive performance, the LR model is recommended as the preferred choice for constructing IAs rupture prediction models, owing to its superior interpretability, computational efficiency, and clinical applicability.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2593/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2593/dss

Funding: This study was supported by the Project of State Key Laboratory of Radiation Medicine and Protection, Soochow University (No. GZK12024025), Scientific Research Project of Jiangsu Provincial Health Commission (No. M2024102), Suzhou Clinical Key Disease Diagnosis and Treatment Technology Special Project (No. LCZX202309), “National Tutor System” Training Program for Health Youth Key Talents in Suzhou (No. Qngg2021006), and the Subject Construction Support Project of The Second Affiliated Hospital of Soochow University (No. XKTJ-HTD2025003).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2593/coif). L.L. and Z.Z. are employees of Shukun Technology Co., Ltd. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The Second Affiliated Hospital of Soochow University (No. JD-LC2024029-I01), and individual consent for this retrospective analysis was waived. The other participating centers were informed of and agreed to this study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Ji Z, Geng JW, Zhai XD, Fu F, Fan XX. Expert consensus on imaging interpretation of intracranial aneurysms. Chin J Cerebrovasc Dis 2021;18:492-504.
2. Neurosurgical Branch of Chinese Medical Association, Cerebrovascular Surgery Branch of Chinese Stroke Association, National Center for Neurological Diseases, National Clinical Research Center for Neurological Diseases. Chinese clinical management guidelines for unruptured intracranial aneurysms (2024 edition). Natl Med J China 2024;104:1918-39.
3. Zhang Z, Li H, Zhou X, Zhong Y, Zhang Y, Deng J, Chen S, Tang Q, Zhang B, Yuan Z, Ding H, Zhang A, Wu Q, Zhang X. Predicting Intracranial Aneurysm Rupture: A Multifactor Analysis Combining Radscore, Morphology, and PHASES Parameters. Acad Radiol 2025;32:359-72. [Crossref] [PubMed]
4. Ou C, Liu J, Qian Y, Chong W, Zhang X, Liu W, Su H, Zhang N, Zhang J, Duan CZ, He X. Rupture Risk Assessment for Cerebral Aneurysm Using Interpretable Machine Learning on Multidimensional Data. Front Neurol 2020;11:570181. [Crossref] [PubMed]
5. Shi Z, Chen GZ, Mao L, Li XL, Zhou CS, Xia S, Zhang YX, Zhang B, Hu B, Lu GM, Zhang LJ. Machine Learning-Based Prediction of Small Intracranial Aneurysm Rupture Status Using CTA-Derived Hemodynamics: A Multicenter Study. AJNR Am J Neuroradiol 2021;42:648-54. [Crossref] [PubMed]
6. Alwalid O, Long X, Xie M, Yang J, Cen C, Liu H, Han P CT. Angiography-Based Radiomics for Classification of Intracranial Aneurysm Rupture. Front Neurol 2021;12:619864. [Crossref] [PubMed]
7. Cai W, Shi D, Gong J, Chen G, Qiao F, Dou X, Li H, Lu K, Yuan S, Sun C, Lan Q. Are Morphologic Parameters Actually Correlated with the Rupture Status of Anterior Communicating Artery Aneurysms? World Neurosurg 2015;84:1278-83. [Crossref] [PubMed]
8. Wei J, Song X, Wei X, Yang Z, Dai L, Wang M, Sun Z, Jin Y, Ma C, Hu C, Xie X, Yang Z, Zhang Y, Lv F, Lu J, Zhu Y, Li Y. Knowledge-Augmented Deep Learning for Segmenting and Detecting Cerebral Aneurysms With CT Angiography: A Multicenter Study. Radiology 2024;312:e233197. [Crossref] [PubMed]
9. Yang L, Gu D, Wei J, Yang C, Rao S, Wang W, Chen C, Ding Y, Tian J, Zeng M. A Radiomics Nomogram for Preoperative Prediction of Microvascular Invasion in Hepatocellular Carcinoma. Liver Cancer 2019;8:373-86. [Crossref] [PubMed]
10. Huang YQ, Liang CH, He L, Tian J, Liang CS, Chen X, Ma ZL, Liu ZY. Development and Validation of a Radiomics Nomogram for Preoperative Prediction of Lymph Node Metastasis in Colorectal Cancer. J Clin Oncol 2016;34:2157-64. Erratum in: J Clin Oncol 2016;34:2436. [Crossref] [PubMed]
11. Greener JG, Kandathil SM, Moffat L, Jones DT. A guide to machine learning for biologists. Nat Rev Mol Cell Biol 2022;23:40-55. [Crossref] [PubMed]
12. Wermer MJ, van der Schaaf IC, Algra A, Rinkel GJ. Risk of rupture of unruptured intracranial aneurysms in relation to patient and aneurysm characteristics: an updated meta-analysis. Stroke 2007;38:1404-10. [Crossref] [PubMed]
13. Shi D, Cai W, Liu R, Jin D, Qiao F, Fan GH. Correlation study between CT angiographic morphological parameters, clinical features and rupture risk of anterior communicating artery aneurysms. J Pract Radiol 2019;35:27-30.
14. Nahed BV, DiLuna ML, Morgan T, Ocal E, Hawkins AA, Ozduman K, Kahle KT, Chamberlain A, Amar AP, Gunel M. Hypertension, age, and location predict rupture of small intracranial aneurysms. Neurosurgery 2005;57:676-83; discussion 676-83. [Crossref] [PubMed]
15. van der Kamp LT, Rinkel GJE, Verbaan D, van den Berg R, Vandertop WP, Murayama Y, et al. Risk of Rupture After Intracranial Aneurysm Growth. JAMA Neurol 2021;78:1228-35. [Crossref] [PubMed]
16. Backes D, Rinkel GJ, Laban KG, Algra A, Vergouwen MD. Patient- and Aneurysm-Specific Risk Factors for Intracranial Aneurysm Growth: A Systematic Review and Meta-Analysis. Stroke 2016;47:951-7. [Crossref] [PubMed]
17. Rinaldo L, Nesvick CL, Rabinstein AA, Lanzino G. Differences in Size Between Unruptured and Ruptured Saccular Intracranial Aneurysms by Location. World Neurosurg 2020;133:e828-34. [Crossref] [PubMed]
18. Grochowski C, Litak J, Kulesza B, Szmygin P, Ziemianek D, Kamieniak P, Szczepanek D, Rola R, Trojanowski T. Size and location correlations with higher rupture risk of intracranial aneurysms. J Clin Neurosci 2018;48:181-4. [Crossref] [PubMed]
19. Rousseau O, Karakachoff M, Gaignard A, Bellanger L, Bijlenga P, Constant Dit Beaufils P, L'Allinec V, Levrier O, Aguettaz P, Desilles JP, Michelozzi C, Marnat G, Vion AC, Loirand G, Desal H, Redon R, Gourraud PA, Bourcier R. Location of intracranial aneurysms is the main factor associated with rupture in the ICAN population. J Neurol Neurosurg Psychiatry 2021;92:122-8. [Crossref] [PubMed]
20. Zeng L, Zhao XY, Wen L, Jing Y, Xu JX, Huang CC, Zhang D, Wang GX. Compare deep learning model and conventional logistic regression model for the identification of unstable saccular intracranial aneurysms in computed tomography angiography. Quant Imaging Med Surg 2024;14:2993-3005. [Crossref] [PubMed]
21. Wang Y, Zhao L, Zhang X, Zheng J, Geng Y, Huang B, Chen T, Qiang J, Liu B, Zhang L, Zhang X. Intracranial aneurysm rupture risk in northern China: a retrospective case-control study. Quant Imaging Med Surg 2024;14:376-85. [Crossref] [PubMed]
22. Xu WD, Shi Z, Hu B, Zhang LJ, Lu GM. Hemodynamics-based analysis of factors associated with aneurysm rupture in different sides of the internal carotid artery. Zhonghua Yi Xue Za Zhi 2021;101:1798-804. [PubMed]
23. Tan J, Zhu H, Huang J, Ouyang HY, Pan X, Zhao Y, Li M. The Association of Morphological Differences of Middle Cerebral Artery Bifurcation and Aneurysm Formation: A Systematic Review and Meta-Analysis. World Neurosurg 2022;167:17-27. [Crossref] [PubMed]
24. Wang GX, Wang S, Liu LL, Gong MF, Zhang D, Yang CY, Wen L. A Simple Scoring Model for Prediction of Rupture Risk of Anterior Communicating Artery Aneurysms. Front Neurol 2019;10:520. [Crossref] [PubMed]
25. Chen Y, Xing H, Lin B, Zhou J, Ding S, Wan J, Yang Y, Pan Y, Zhao B. Morphological risk model assessing anterior communicating artery aneurysm rupture: Development and validation. Clin Neurol Neurosurg 2020;197:106158. [Crossref] [PubMed]
26. Ou C, Chong W, Duan CZ, Zhang X, Morgan M, Qian Y. A preliminary investigation of radiomics differences between ruptured and unruptured intracranial aneurysms. Eur Radiol 2021;31:2716-25. [Crossref] [PubMed]
27. Liu Q, Jiang P, Jiang Y, Ge H, Li S, Jin H, Li Y. Prediction of Aneurysm Stability Using a Machine Learning Model Based on PyRadiomics-Derived Morphological Features. Stroke 2019;50:2314-21. [Crossref] [PubMed]
28. Zhu D, Chen Y, Zheng K, Chen C, Li Q, Zhou J, Jia X, Xia N, Wang H, Lin B, Ni Y, Pang P, Yang Y. Classifying Ruptured Middle Cerebral Artery Aneurysms With a Machine Learning Based, Radiomics-Morphological Model: A Multicentral Study. Front Neurosci 2021;15:721268. [Crossref] [PubMed]
29. Sohrabi-Ashlaghi A, Azizi N, Abbastabar H, Shakiba M, Zebardast J, Firouznia K. Accuracy of radiomics-Based models in distinguishing between ruptured and unruptured intracranial aneurysms: A systematic review and meta-Analysis. Eur J Radiol 2024;181:111739. [Crossref] [PubMed]
30. Ma X, Yang Y, Liu D, Zhou Y, Jia W. Demographic and morphological characteristics associated with rupture status of anterior communicating artery aneurysms. Neurosurg Rev 2020;43:589-95. [Crossref] [PubMed]
31. Lindgren AE, Koivisto T, Björkman J, von Und Zu Fraunberg M, Helin K, Jääskeläinen JE, Frösen J. Irregular Shape of Intracranial Aneurysm Indicates Rupture Risk Irrespective of Size in a Population-Based Cohort. Stroke 2016;47:1219-26. [Crossref] [PubMed]
32. Detmer FJ, Lückehe D, Mut F, Slawski M, Hirsch S, Bijlenga P, von Voigt G, Cebral JR. Comparison of statistical learning approaches for cerebral aneurysm rupture assessment. Int J Comput Assist Radiol Surg 2020;15:141-50. [Crossref] [PubMed]
33. Fujimura S, Yamanaka Y, Takao H, Ishibashi T, Otani K, Karagiozov K, Fukudome K, Yamamoto M, Murayama Y. Hemodynamic and morphological differences in cerebral aneurysms between before and after rupture. J Neurosurg 2024;140:774-82. [Crossref] [PubMed]
34. Skodvin TØ, Johnsen LH, Gjertsen Ø, Isaksen JG, Sorteberg A. Cerebral Aneurysm Morphology Before and After Rupture: Nationwide Case Series of 29 Aneurysms. Stroke 2017;48:880-6. [Crossref] [PubMed]
35. Schneiders JJ, Marquering HA, van den Berg R, VanBavel E, Velthuis B, Rinkel GJ, Majoie CB. Rupture-associated changes of cerebral aneurysm geometry: high-resolution 3D imaging before and after rupture. AJNR Am J Neuroradiol 2014;35:1358-62. [Crossref] [PubMed]