Radiomics Signatures of Carotid Plaque on Computed Tomography Angiography

This retrospective study was approved by the local institutional review committee (No. Y [2022]114), and the requirement for informed consent was waived. We included 1063 consecutive patients with clinically suspected stroke who underwent head and neck CTA and brain MRI within 1 month between October 2019 and July 2022. The exclusion criteria were as follows: (1) insufficient clinical data; (2) negative findings on carotid CTA; (3) presence of cerebral hemorrhage, tumor, trauma or previous brain surgery; (4) posterior circulation stroke; (5) suspected cardioembolic sources; (6) carotid artery dissection; (7) accompanied by intracranial vessel diseases, such as atherosclerotic stenosis, aneurysm or moyamoya disease; (8) previous carotid stenting and endarterectomy; (9) insufficient image quality for plaque analysis and radiomics extraction (see Fig. 1 for the flow chart of patient selection).

Demographic, clinical and laboratory data were collected from all patients, including age, sex, smoking status, hypertension, diabetes mellitus, hyperlipidemia, coronary artery disease, medication history, total cholesterol, triglycerides, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, lipoprotein(a) and homocysteine (Hcy).

Head and neck CTA was performed using a 256-slice CT scanner (Brilliance iCT; Philips Medical Systems, Cleveland, OH, USA). The aortic arch to the skull vertex was covered by CTA examination in the helical scanning mode. The scanning parameters were as follows: detector collimation, 128 × 0.625 mm; tube voltage, 120 kV; tube current, 250 mA; rotation time, 0.33 s; pitch, 0.763; section increment, 0.45 mm; and slice thickness, 0.9 mm. Approximately 50–60 mL of nonionic iodine contrast agent was injected at a flow rate of 4.0–5.0 mL/s via a high-pressure injector, followed by a 20–30 mL flush of saline at the same rate.

The MRI examinations were performed on a 3.0T MR scanner (Discovery 750; GE Healthcare, Chicago, IL, USA) with a 20-channel head coil using standardized protocols, including T1 and T2 weighted imaging, T2 fluid-attenuated inversion recovery (FLAIR) and diffusion weighted imaging (DWI). Ischemic lesions were assessed on DWI with the following parameters: TR 4000 ms, TE 68 ms, FOV 24 cm×24 cm, slice thickness 6 mm, matrix 128 × 128, intersection gap 1.2 mm, number of excitations 1, b‑values of 0 and 1000 s/mm².

Carotid plaque characteristics were assessed on CTA using GE Healthcare Advantage Workstation (AW4.4, USA). Carotid stenosis, plaque features and carotid artery tortuosity were measured and evaluated as follows: (1) the degree of carotid stenosis was defined as mild (< 50%) and moderate to severe (50–99.9%) stenosis using the North American Symptomatic Carotid Endarterectomy Trial (NASCET) criteria [21]; (2) the maximum plaque thickness: measured on the cross-section of the plaque perpendicular to the long axis of the vessel, (3) the soft plaque thickness: a plaque with a mean attenuation of 40–50 HU was considered soft [5]; (4) plaque ulceration: defined as at least 1 mm of contrast agent entering the plaque in any one plane [22]; (5) carotid rim sign: a positive carotid rim sign was defined as adventitial calcification (< 2 mm thickness) with interior soft plaque (> 2 mm thickness), whereas a negative carotid rim sign was defined as adventitial calcification (< 2 mm thickness) with little interior soft plaque (< 2 mm thickness) ([23]; Fig. 2); (6) positive remodelling (PR): characterized as an outside vessel diameter that was 10% larger than the mean diameter of the segments that were both proximal and distal to the plaque [24] and (7) common carotid tortuosity index (CCTI) and internal carotid tortuosity index (ICTI): calculated using the following formula: ([actual distance/linear distance] − 1) × 100 [25]. The actual distance of the vessel was measured along the arterial center line from the proximal to the distal point in the curved projection reformation. The linear distance of the vessel was determined as the shortest distance between the same two points in the three-dimensional volume rendering image. Proximal vessel points were uniformly placed at the arterial origin. The common carotid artery was placed distally at the carotid bifurcation, and the ICA was placed distally at the entrance of the outer carotid canal (Fig. 2).

Patients were divided into symptomatic and asymptomatic groups according to the following criteria: (1) clinical symptoms continuing for at least 24 h within 2 weeks before imaging scans; and (2) acute/subacute stroke findings on DWI in the ipsilateral carotid artery territory [26, 27].

Patient data were anonymized and transmitted to The Deepwise Multimodal Research Platform v2.0 (Beijing Deepwise & League of PHD Technology Co., Ltd, Beijing, China, https://​keyan.​deepwise.​com), which can be used for labeling, region of interest (ROI) feature extraction, and model development [20, 28]. In the symptomatic group, we segmented the carotid plaque located ipsilateral to the infarct. For patients with bilateral plaques in the asymptomatic group, we segmented unilateral plaque with a larger volume. For patients with plaques involving both the internal and external carotid arteries, we only segmented the internal carotid plaques. A radiologist visually assessed the outer border of the arterial wall and segmented the plaque boundaries with an appropriate center level and window width.

We selected clinical and conventional plaque characteristics with statistical differences in univariate and multivariable logistic regression analyses and built the conventional model. For radiomics model development, we first preserved radiomic features with intraobserver and interobserver ICC > 0.8 and eliminated redundant features with linear correlation coefficient values (threshold = 0.7). We then used the analysis of variance F‑value, mutual information, and linear models penalized with the L1 norm algorithm to form radiomics signature. Finally, a combined model was created by integrating selected conventional characteristics and radiomics signature. All models were trained using the fivefold cross-validation method. The workflow of the model development is shown in Fig. 3.

In total, 167 patients were included in the final analysis. Patients were mainly older (66.2 ± 7.7 years; 78.4% male), among whom smoking (54.5%) and hypertension (68.9%) accounted for a higher proportion. The clinical characteristics of patients in the symptomatic and asymptomatic groups are shown in Table 1. Smoking was more prevalent in the symptomatic group than in the asymptomatic group (65.7% vs. 46.4%, p = 0.013). Total cholesterol (4.7 vs. 4.2 mmol/L, p = 0.022), lipoprotein(a) (231.4 vs. 134.2 mg/L, p < 0.002), and Hcy (13.9 vs. 11.0 µmol/L, p < 0.001) were higher in the symptomatic group than the asymptomatic group. Multivariate logistic regression analysis revealed that Hcy (odds ratio, OR, 1.057; 95% confidence interval, CI, 1.001–1.116) was the only independent risk factor associated with symptomatic plaques (Table 2).

Univariate analysis showed that the degree of carotid artery stenosis (OR 1.02; 95% CI 1.01–1.04), total plaque thickness (OR 1.44; 95% CI 1.16–1.79), soft plaque thickness (OR 1.64; 95% CI 1.31–2.04), plaque length (OR 1.06; 95% CI 1.01–1.10), plaque ulceration (OR 9.00; 95% CI 3.22–25.20), carotid rim sign (OR 6.57; 95% CI 2.75–15.7), and PR (OR 2.01; 95% CI 1.08–3.76) were related to plaque vulnerability. Further multivariate logistic regression analysis revealed that plaque ulceration (OR 6.106; 95% CI 1.933–19.287) and carotid rim sign (OR 3.285; 95% CI 1.203–8.969) were independent predictors of symptomatic plaques (Table 2). Intraobserver and interobserver reproducibility of conventional imaging features assessment was excellent (supplementary materials Table S1).

A total of 2107 radiomic features were extracted and normalized using the Z‑score from each ROI. The ICC for calculating radiomic features is summarized in the supplementary material (Fig. S1). First, radiomic features with ICC ≤ 0.8 were removed by interobserver and intraobserver agreement, and the remaining 1916 features were included in the analysis. Another 112 features were removed because of low discrimination. Then 1760 features were removed according to the linear correlation value (threshold = 0.7). A mutual information algorithm was used to analyze all remaining features. Finally, six significant radiomic features were retained to construct the radiomics model by using the logistic regression algorithm provided by the Deepwise Multimodal Research Platform [20], including wavelet-HLL_gldm_GrayLevelNonUniformity, wavelet-LLH_firstorder_Mean, wavelet-HLL_glcm_Imc2, log-sigma-2-0-mm-3D_gldm_LargeDependenceLowGrayLevelEmphasis, square_glcm_InverseVariance, and the exponential_ngtdm_Strength.

We developed three formulas based on independent risk factors and radiomics signature to calculate each patient’s risk score. They were presented as follows:

Table 3 shows the discriminative power of the three models. The radiomics model demonstrated a good diagnostic performance in the training (area under the curve, AUC 0.840; 95% CI 0.781–0.898) and validation cohorts (AUC 0.797; 95% CI 0.731–0.863). The conventional model had lower discrimination for symptomatic plaques in the training (AUC 0.789; 95% CI 0.715–0.863) and validation cohorts (AUC 0.767; 95% CI 0.689–0.845). The AUCs of the integrated model were the highest, with 0.856 (95% CI 0.800–0.913) and 0.832 (95% CI: 0.770–0.894) for the training and validation sets, respectively (Fig. 4). DeLong’s test showed that the integrated models significantly improved model performance compared to the conventional model in training (AUC 0.856 vs. 0.789, p = 0.005) and validation cohorts (AUC 0.832 vs. 0.767, p = 0.011). Although the diagnostic efficiency of the integrated model relatively outperformed the radiomics model, there was no significant difference in the training (AUC 0.856 vs. 0.840; p = 0.688) and validation cohorts (AUC 0.832 vs. 0.797; p = 0.452).

On the basis of the combined model, a nomogram was constructed that incorporated three selected conventional features (Hcy, plaque ulceration, and carotid rim sign) and radiomics score for predicting plaque vulnerability (Fig. 5). The value of each of these variables was given a score on the point scale axis. A total score could be easily calculated by adding each single score to the total point scale to estimate the risk of plaque vulnerability.

The calibration curve of the combined model demonstrated a good fit between predicted and actual possibility of ischemic stroke in the training and validation cohorts (Fig. 6). The decision curves compare the net benefits at various threshold probabilities in the training and validation sets to demonstrate the clinical utility of the prediction model. The integrated model had an excellent overall net benefit when the risk threshold range was 0.15–0.82 (Fig. 7).