Consecutive patients imaged at Sichuan Provincial People's Hospital from November 2015 to December 2019 who were diagnosed as having intracranial aneurysms on CTA were retrospectively reviewed. Among these patients, the study population was selected based on the following inclusion criteria: (1) confirmed diagnosis of one or more unruptured saccular cerebral aneurysms at baseline by CTA (positive diagnosis in the clinical record by experienced radiologist); (2) aneurysm maximum diameter >2 mm and lesion morphology that is differentiable from infundibulum; (3) image quality allowed detailed depiction of aneurysm morphology and parent vessels, and was sufficient for the segmentation without substantial artifacts. Exclusion criteria: (1) fusiform, dissecting, traumatic, or infectious aneurysms; (2) the presence of concomitant cerebral vascular diseases, such as high-flow arteriovenous malformation, Moyamoya disease, arteriovenous fistula, dissection, vasospasm and inflammatory vasculopathy; (3) acute aneurysm rupture with presence of subarachnoid hemorrhage; (4) aneurysm endovascular treatment prior to CTA imaging.

Included aneurysms were divided into two groups according to their stability, following the criteria of a previous study (12) and supported by multiple studies (20, 21). The positive group (unstable aneurysms, considered with high risk of rupture) consists of 3 subgroups. The first subgroup is aneurysms with a near-term rupture event that occurred after CTA image acquisition (within 4 days). This category included patients who presented with sentinel headaches, negative non-contrast CT head, and presence of aneurysm on CTA. Aneurysm rupture was subsequently diagnosed on follow-up non-contrast CT head, with presence of hyperdense subarachnoid hemorrhage. We did not include ruptured aneurysms because aneurysm morphology changes after rupture (14). The second subgroup in the positive group is growing aneurysms. Growing aneurysms were identified on sequential angiographic imaging where the maximum diameter increased by more than 1 mm between two imaging exams according to the ELAPSS criteria (22). The third subgroup is symptomatic aneurysms, defined as patients with aneurysms causing compressive symptoms (ex. oculomotor nerve compression confirmed intraoperatively) (12). Other less specific symptoms including dizziness and headaches which could not be definitively attributed to aneurysm were not included in the symptomatic group. The negative group (stable aneurysms, considered to have low rupture risk) included incidentally discovered asymptomatic aneurysms or those undergoing follow-up without rupture or growth. When multiple aneurysms were identified, those that met the inclusion and exclusion criteria were included in the analysis, and each aneurysm was assessed and classified into the positive or negative group, respectively. For instance, if a patient with multiple aneurysms had near-term SAH and underwent intervention, the ruptured aneurysm was classified as previously unstable (at the time of the prior CTA), and the remaining unruptured aneurysms were classified as stable.

A flowchart of the patients' inclusion and exclusion process was shown in Figure 1.

Demographic and clinical information were collected from the medical record which included age, sex, history of hypertension, hyperlipidemia, diabetes, smoking and alcohol use. Hypertension is diagnosed if the mean blood pressure at rest is ≥130 mmHg systolic or ≥80 mmHg diastolic. Hyperlipidemia is generally diagnosed at a fasting serum (or plasma) TG (triglyceride) level of ≥150 mg/dL (1.7 mmol/L). FPG (fasting plasma glucose) values ≥126 mg/dL (7.0 mmol/L) can be diagnosed as diabetes, where fasting is defined as no caloric intake for at least 8 h.

The imaging protocol included standard CT angiography of the head or head and neck acquired on SOMATOM Definition AS + (Siemens Healthineers, Erlangen, Germany) from Sichuan Provincial People's Hospital. The scans employed the following parameters: kV, 100; mA, 250–300; section thickness, 0.75 mm; reconstruction interval, 0.5 mm; scanning direction, from the second cervical vertebra to the vertex. Iohexol non-ionic contrast agent (Shanghai General Electric Pharmaceutical Co., Ltd, Shanghai, China) was injected through elbow vein with double syringe at an injection rate of 4.5 ml/s. An automatic fluoroscopic bolus–triggered system was set to determine the timing of data acquisition. The scanning was started when contrast agent concentration reached threshold levels at the C2-C3 level of the internal carotid artery.

Traditional morphological features identified from the included CTA images were as follows:

Dmax (Maximum diameter) was defined as the largest distance between any two points along the surface of the aneurysm.

Shape was considered as regular or irregular (aneurysms with daughter sacs, multiple lobes, or other types of wall protrusions).

Dn (Neck diameter) was computed as twice the average distance from the neck centroid to the neck border.

Height (Aneurysm maximum height) was defined as the largest distance between the center of the neck (geometric center) and the aneurysm surface.

Dv (Parent vessel diameter) was computed as twice the average distance from the parent vessel cross section centroid to the parent vessel plane border.

AR (Aspect ratio) was defined as the ratio between Maximum height and Neck diameter.

SR (Size ratio) was defined as the ratio between Maximum height and Parent vessel diameter (23).

Location of aneurysms was divided into 4 categories: intracranial intradural carotid artery and posterior communicating artery, middle cerebral artery, anterior cerebral artery circulation (anterior cerebral artery and anterior communicating artery), and posterior circulation (vertebral artery, basilar artery, and posterior cerebral artery).

All features of IAs were evaluated and measured by two raters (R.L. and P.Z.) and confirmed by one senior radiologist with 10 years of clinical imaging experience (Y.W.). Two raters measured all samples, respectively. The inter-rater reliabilities of the traditional morphological features were performed by evaluating intra-class correlation.

Accounting for the variation in slice thickness, before feature extraction, B-spline algorithm was used to re-slice the images to standard template of voxels with 1 × 1 × 1 mm size. Radiomics software (Prototype, Siemens-healthineers, Erlangen, Germany) was used for identifying Regions of interest (ROIs). The two raters drew manual ROIs slice-by-slice on the original sagittal, reconstructed axial, and coronal CTA images, outlining the aneurysm boundary to cover the entire lesion volume. Each rater drew half the samples in order independently and an additional 20% of the patients for reliability testing. ROIs included as few peripheral structures (adjacent blood vessels or soft tissues) as possible. After the ROI segmentation was completed, Radiomics features were extracted automatically by using the calculation function of the software, and the features were extracted according to the feature guidelines of the Image Biomarker Standardization Initiative (IBSI) (24).

First-order statistics, shape (2D), shape (3D), gray-level co-occurrence matrix (GLCM), gray-level size-zone matrix (GLSZM), gray-level run-length matrix (GLRLM), neighborhood gray-tone difference matrix (NGTDM), and neighboring gray-level dependence matrix (GLDM) were the eight feature groups that were extracted. Advanced filters including Laplacian of Gaussian (LoG; sigma, 2.0 and 3.0 mm), wavelet decompositions with all possible combinations of high- (H) or low- (L) pass filter in each of the three dimensions (HHH, HHL, HLH, LHH, LLL, LLH, LHL, HLL), and local binary pattern (LBP; level, 2; radius, 1.00) were also applied. Definitions and calculations of the Radiomics features used in this study can be found in the PyRadiomics documentation (24, 25).

Stratified random sampling technique was applied to create two random samples of training and independent testing sets within studied classes. Given the relatively low percentage of unstable aneurysms (36/254 = 14.2%), we used a ratio of 1:1 for samples size allocation for two sets. If the training vs. test ratio was larger, such as 6:4 or even 7:3, the sensitivity and specificity of the model would swing disproportionately by the misjudgment of one single positive case in the test set. In consideration of this issue, increasing the proportion of test sets was suggested from a statistic perspective (26).

The training set was used to train the model based on least absolute shrinkage and selection operator (LASSO) regression model with tuning parameter of lambda. Since there was class imbalance that might affect the model tuning, synthetic minority over-sampling technique (SMOTE) was applied to the training sets before formal model training. The lambda was tuned across multiple values between 0.01 and 2. The optimal value of lambda was decided when the model gained the highest AUC based on 5 repeats 10-folds cross-validation. Subsequently, the importance of the features based on this model with optimal lambda were ranked. Next the conventional logistic multivariate regression model was fitted. To ensure both the inferential and prediction performance of this statistical model, we only enrolled a limited number of features based on the above mentioned importance rank list. The number was decided by obeying the rule suggested by Green (27), that N ≥ 50 + 8 m, where m is the number of features and N is the number of samples.

The model performance was evaluated in training and testing sets, respectively. These two apparent performances were shown in terms of receiver operator curves. Sensitivity, specificity, and accuracy were also illustrated using a confusion matrix. In addition, to allow for the estimation of the model generalizability, bootstrapping resampling methods were applied to SMOTE training sets to generate 5 samples for model evaluation.

Three basic prediction models were built using selected features to predict aneurysm stability. Model A was constructed using traditional morphological parameters measured directly from images. Model B was based on the selected Radiomics features calculated by the software. Among different categories of Radiomics features, morphological features, especially advanced ones, such as flatness and elongation, have gained considerable attention in the literature as potential novel biomarkers for aneurysm stability (18). In order to specifically assess the performance of these features, Model C was constructed with only Radiomics features related to morphology. It has been argued that the utility of grayscale image characteristics and texture parameters in Radiomics features are limited in assessing intracranial aneurysms, and morphological parameters are more meaningful (18). Therefore, we separately analyzed and verified the morphological parameters in Radiomics by Model C in addition to the Radiomics model, Model B. Integrated models of traditional and Radiomics features were also built (model A+B, model A+C).

Model A, model B, model C and integrated models were evaluated and compared by calculating the AUC on both the training and test sets.

The Shapiro-wilk test was used to test the normality of continuous variables. Continuous variables were summarized as means ± standard deviation if normally distributed, or median and interquartile ranges (IQRs) if otherwise. Categorical variables were presented as percentages. Fisher exact test or chi-square test was used for comparisons between groups for categorical variables, and Mann-Whitney U-test was used for continuous variables. Univariate and multivariate analyses were performed sequentially. Binary logistic regression was performed with a backward stepwise method. Correlation of variables was assessed by Pearson Correlation Coefficient for normally distributed continuous variables, and by Spearman's Rank-Order Correlation otherwise. The statistical significance was defined as p < 0.05 (two-tailed). To assess inter-rater variability, we used the intra-class correlation coefficient (ICC) for each feature (two-way random effects, single rater/measurement, absolute agreement, with the value higher than 0.9 regarded as excellent). For statistical analyses, SPSS (Statistics Version 26, Chicago, IL, United States) was used.

A total of 227 patients [149 (65.6%) females] with 254 intracranial aneurysms met the criteria and were included. There were 31 (13.7%) patients with 36 (14.2%) unstable aneurysms in the positive group. Fifteen of the 35 unstable aneurysms had a near-term rupture event, 16 had growth in sequential imaging follow-up (median, 6.5 months; range, 0.5–35 months), and 5 had compressive symptoms, respectively. The demographic, clinical, and imaging characteristics of the unstable and stable groups are shown in Table 1. No significant differences were present in patients' age, sex, history of hypertension, hyperlipidemia, diabetes, smoking and alcohol use between the two groups (all p > 0.05). After splitting, there were 120 patients allocated to the training set, with the remaining cases placed in the testing set. Considering the training set sample size, we could enroll 8 features maximum for each model.

Inter-rater reliability for traditional morphology measurements was found to be good (mean ICC = 0.969; range 0.952–0.987).

A total of 1,691 Radiomics features were extracted and included in the intra-class correlation evaluation. Features with ICC of <0.9 were excluded. A total of 460 features were excluded for unsatisfactory inter-rater agreement. Ultimately, 1,231 features were selected for further analysis. The overall inter-rater agreement of the 1,231 features was excellent (mean ICC = 0.966; range 0.900–1.000).

The performance of the integrated models of traditional and Radiomics features as evaluated by ROC are shown in Figure 6. The AUC value of the integrated model A+B was 0.868, which was slightly higher than any single model, but not statistically significant. The integrated model A+C failed to improve diagnostic performance in AUC compared to model A.

Flatness was the only advanced morphological feature among the selected features. The relationship between flatness and traditional morphological features was evaluated as follows: among traditional morphological features, including Height, Dmax, Dn, Dv, AR, and SR, Flatness correlated best with Height (p < 0.01, R² = 0.19), followed by SR (p < 0.05, R² = 0.13), as shown in Figure 7. The correlation between flatness and each traditional morphological feature is shown in Table 4.

Our results show that the model performance based on traditional morphological features is still good (AUCs of 0.888 and 0.818, respectively in training and testing sets). By comparison, the Model A score had little overlap between the two groups, as shown in the box plots (Figure 2). Features defined by ratios such as AR were shown as strong factors related to aneurysm rupture [also supported by our previous large-sample study of a similar population (30)], which may not be fully presented in the definition of Radiomics features. A recent study using machine learning to classify intracranial aneurysm rupture status on CTA also screened out mainly traditional morphological features, such as irregular shape and size ratio as the most important discriminators (31). Another study also revealed that Radiomics were not obviously superior to traditional morphological features (18), which is consistent with our study. Although traditional morphological features and Radiomics features can be integrated, only the model integrating traditional features and overall Radiomics features (model A+B) showed a marginal improvement in performance. The use of Radiomics-derived morphological features alone (Model C), however, showed poor performance.

This study has several novel features compared to previous studies. The study by Liu et al. was based on DSA (12), which is invasive and not used for screening and re-examination. CTA was used in this study as it is the most frequently employed imaging modality for aneurysm evaluation. Previous studies only assessed the region of interest from a single image slice (32), while we delineated the region of interest from all image slices with a slice thickness of 0.75 mm, and extracted three-dimensional aneurysm volumes as the basis of radiomic and morphological characteristic analysis.

Two previous Radiomics studies based on CTA images aimed to distinguish ruptured and unruptured aneurysms, considering already ruptured aneurysms as unstable ones (18, 19). Aneurysm morphology, however, may change after rupture (14), and the identified biomarkers and corresponding conclusions may not be directly applied to predict the risk in the group of unruptured aneurysms. In this study, more comprehensive and representative criteria of aneurysm instability were used, and all the included aneurysms were considered to have not yet ruptured at the time of CTA imaging. The conclusions can therefore be more directly interpreted and used clinically to predict rupture risk.

As reported, the proportion of aneurysm growth in follow-up studies was relatively low (mostly reported around 5%, one study reported 12% in 2 years) (33–35). Aneurysms at near-rupture status are difficult to obtain clinically. In this study, categorized 14.2% (N = 36) aneurysms into the unstable group, which reduced the limitations of data imbalance and facilitated model construction. Radiomics calculation is a fast and automatic process using one-station software. Once the structures of interest have been semi-automatically delineated, the selected Radiomics features and the calculation of Rad-score can be readily employed in clinical practice with further investigation and validation.