Ji-Yan Nie ^1,2* Wen-Xi Chen ^1,2* Zhi Zhu ^1,2* Ming-Yu Zhang ^1,2* Yu-Jin Zheng ^1* Qing-de Wu ^1*

• ¹ Department of Radiology, Shunde Hospital of Guangzhou University of Chinese Medicine, Foshan, China
• ² Graduate School, Guangzhou University of Chinese medicine, Guangzhou, China

Carotid atherosclerotic disease is the main cause of ischemic stroke, accounting for about 34% of ischemic stroke (1). The guidelines for the prevention and treatment of Stroke in China 2021 recommend carotid endarterectomy or carotid artery stenting for patients with more than 50% carotid artery stenosis to prevent stroke. However, the degree of carotid artery stenosis does not completely match the occurrence of stroke (2), and there is currently a lack of objective indicators to assess the risk of stroke in carotid plaque. Head and neck computed tomography angiography (CTA) is the first line non-invasive imaging method for carotid atherosclerosis (3). Radiomics analysis of carotid plaques based on CTA has made some progress in identifying carotid plaques at high risk of stroke. However, automatic segmentation of carotid plaques is challenging due to the complex composition of plaques and the limited number of pixels in CTA images. As a consequence, the radiomic signature (RS) model derived from these segmentations often exhibits low performance and lacks universality (4). Vascular inflammation can drive atherosclerotic plaque rupture and thrombosis, leading to the occurrence of adverse cardiovascular and cerebrovascular events (5). A considerable body of recent research (6)(7)(8)(9)(10) has demonstrated that perivascular adipose tissue (PVAT) can be automatically segmented by applying a threshold range of -190 to -30HU on CTA, enabling the monitoring of vascular inflammation and identification of symptomatic plaques. Numerous studies (11)(12)(13)(14) have also indicated that the pericoronary adipose tissue RS model exhibits excellent performance in identifying and predicting symptomatic plaques; however, there is limited literature available regarding carotid artery investigations.In this study, we used radiomics analysis combined with machine learning methods to establish a RS model based on the PVAT of carotid plaques combined with traditional patient characteristics, and investigated its performance in distinguishing symptomatic and asymptomatic carotid plaques. Materials and methods This was a retrospective study involving patients who underwent head and neck CTA at our hospital from April 2021 through February 2023 (Figure 1). All patients were divided into symptomatic group and asymptomatic group according to whether they had clinical symptoms within 2 weeks before CTA examination and/or whether head MRI showed acute/subacute stroke (15). Clinical symptoms included classic TIA (transient ischemic attack) and anterior circulation (carotid territory) ischemic stroke, as well as monocular symptoms ipsilateral (16) to the carotid plaque (amaurosis or retinal artery occlusion). Classic TIA is defined as an abnormal focal neurological deficit lasting less than 24 hours. Complete ischemic stroke presents with the sudden onset of a focal neurologic deficit lasting > 24 hours (17). The patient's age, gender, body mass index (BMI), history of hypertension, diabetes, hyperlipidemia, smoking history, history of antihypertensive drugs, and history of antiplatelet drugs were collected. Head and neck CTA was performed using a third-generation dual source CT (Somatom Force, Siemens). The patient was placed in a supine position with head advanced and calm breathing. The scanning direction was the foot head direction, and the scanning range was from the level of the sternal Angle to the skull dome. 50 ml of ioversol (Bayer, Germany, iodine concentration 370mg/ml) was injected via the cubital vein with a high pressure syringe at a rate of 5ml/s, and 40ml of normal saline was injected at the same flow rate. The ROI was drawn at the descending aortic arch using contrast agent tracking technology. The trigger threshold was 100HU, and the scan was delayed for 3-4s after trigger. The tube voltage was 90-100 KVp, and the tube current was adaptive. All CTA data were transferred to head and neck CTA AI system (Shukun Technology, Beijing, China) for plaque localization and analysis on curved planar reconstruction images. The symptomatic group selected the narrowest carotid plaque on the symptomatic side, and the asymptomatic group selected the narrowest carotid plaque. According to the location of the plaque, the plaque was divided into left carotid artery plaque and right carotid artery plaque. The degree of plaque stenosis was automatically calculated.Plaque thickness was measured as the maximum axial size of the plaque on a single axial slice, representing its maximum thickness. Plaque length was defined as the distance from the origin of the plaque to the distal end. The remodeling index was calculated by averaging the maximum external vessel diameter of the plaque over the normal diameter of the proximal and distal regions.Plaques were classified into three types based on the presence or absence of calcification: calcified plaque, non-calcified plaque, and mixed plaque. The presence of plaque ulceration was identified by the spread of contrast agent deep into the plaque on multiple slices from different imaging perspectives. High-risk plaque is defined as having two or more of the following features: positive remodeling index > 1.1, punctate calcification (with a diameter < 3mm, occupying < 1/4 of the lumen's diameter, and a CT value > 130HU), low-density plaque (a non-calcified plaque with a CT value < 30HU and an area of 1mm² within the plaque), and the napkin ring sign (a contrast agent ring encircling a low-density plaque component, along with contrast agent in the surrounding vascular lumen). ROI segmentation of the PVAT of extracranial carotid plaques was performed using perivascular fat analysis software (Shukun Technology, Beijing, China). The measurement was centered on the carotid bifurcation, extending 2 cm in the superior and inferior directions for a total length of 4 cm. The PVAT width equivalent to the diameter of the carotid artery beyond the outer wall of the carotid artery vessel. The software automatically segmented adipose tissue with an attenuation value of -190 HU to -30 HU along the target length and a width of the carotid artery vessel (25-26)(Figure 2). The Fat Attenuation Index (FAI) surrounding atherosclerotic plaques was assessed using specialized perivascular fat analysis software (Shukun Technology, Beijing, China). The length of "Stenosis FAI" is measured on the narrowest cross-sectional slice of the plaque, while the length of "Stenosis range FAI" is measured along the entire extent of the plaque, from its origin to the distal end. Both FAI measurements have a width equivalent to the diameter of the carotid artery beyond the outer wall of the carotid artery vessel. The software automatically segmented adipose tissue with an attenuation value of -190 HU to -30 HU along the target length and a width of the carotid artery vessel, following which the software automatically computes the average density of the perivascular fat encompassing the plaque (Figure 2). ROI of all plaque PVAT was imported into Shukun AI Scientific Research Platform (Beijing, China) for RS extraction. A total of 1874 RS were extracted from the ROI of each plaque PVAT. These included 360 first-order features, 14 shape features, 480 gray level co-occurrence matrix (GLCM), 280 gray level dependence matrix (GLDM), 320 gray level run length matrix (GLRLM), 320 gray level size zone matrix (GLSZM), and 100 neighborhood gray tone difference matrix (NGTDM). All The data analysis was performed using SPSS The machine learning algorithm parameters used in both models were identical to those in the Model 1. Area Under the Curve (AUC) was used to evaluate the ability of the two groups of models to identify symptomatic plaques. DeLong test was used to compare the difference between AUCs. P<0.05 was considered statistically significant (Figure 3). This study included a total of 203 patients, with an average age of 71.87±9.63 years and a total of 115 males. Among them, there were 71 cases in the symptomatic group and 132 cases in the asymptomatic group.In the multivariate logistic regression analysis, it was found that the proportion of positive remodeling in the symptomatic group was higher than that in the asymptomatic group (97.2% vs 84.8%, P=0.017). Additionally, the proportion of statin use in the symptomatic group was significantly lower than that in the asymptomatic group (15.5% vs 47%, P < 0.001).Other factors such as age, gender, BMI, history of hypertension, diabetes mellitus, hyperlipidemia, smoking history, history of antihypertensive drugs, history of antiplatelet drugs, plaque location, degree of plaque stenosis, plaque length, plaque thickness, remodeling index, FAI at the most stenosis of the plaque, FAI within the stenosis of the plaque, conformal remodeling, low-density plaque, punctate calcification, napkin ring sign, and high-risk plaque distribution, plaque type, and plaque ulcer did not show statistically significant differences in the multivariate regression analysis (P > 0.05) (Table 1). The RS model showed the highest diagnostic performance in identifying symptomatic plaques within the Bagging Decision Tree model, achieving an AUC of 0.837 (95%CI: 0.775, 0.899) in the training set and an AUC of 0.834 (95%CI: 0.685, 0.982) in the testing set. These results were significantly better than the performances of the XGBOOST, Random Forest, SVM, and QDA models (P < 0.05) (Figure 4) (Table 2). The figure 5 depicts the diagnostic performance of traditional feature models and the RS model in identifying symptomatic plaques across different sets of data. In the training set, the traditional feature model achieved the AUC of 0.725 (95%CI: 0.695, 0.791), while in the testing set, the AUC was 0.593 (95%CI: 0.438, 0.749). Upon incorporating the RS model into the traditional feature model, the AUC in the training set improved to 0.831 (95%CI: 0.765, 0.896), and in the testing set, it reached 0.82 (95%CI: 0.675, 0.965).Through the Delong test, it was determined that the combination of the RS model with the traditional feature model yielded a significantly higher AUC for distinguishing symptomatic plaques compared to using the traditional model alone (AUC: 0.82 vs AUC: 0.593; Z = 2.822, P = 0.0048). Furthermore, when used independently, the RS model demonstrated a superior AUC in differentiating symptomatic plaques compared to the traditional model (AUC: 0.834 vs AUC: 0.593; Z = 2.114, P = 0.0345). This study confirms that the RS model, based on carotid PVAT, has demonstrated significant improvement over the current traditional models in distinguishing symptomatic plaques. The RS model, relying on carotid PVAT, exhibited a higher AUC in the discrimination of symptomatic plaques (AUC: 0.834; 95% CI: 0.685, 0.982), compared to the traditional model (AUC: 0.593; 95% CI: 0.438, 0.749).CT-based radiomics has been shown to be able to accurately classify diseases by extracting a large number of quantitative radiomics features that are invisible to the human eye (27) A large study showed(29) that long-term achievement of low LDL-C levels, as low as less than 20 mg per deciliter (<0.5 mmol per liter), was associated with a reduced risk of cardiovascular outcomes without significant safety concerns in patients with atherosclerotic cardiovascular and cerebrovascular disease. The lower proportion of statin use in the symptomatic group than in the asymptomatic group in this study may be due to the higher incidence of ischemic stroke or TIA in patients who do not receive statin therapy.There was no significant difference in FAI at the narrowest point of the plaque and within the stenosis of the plaque between the two groups, which may be related to the fact that most of the patients in this study were elderly with an average age of 71.87±9.63 years. FAI is used to dynamically monitor vascular inflammation ( 5) by measuring the mean density of adipose tissue on CT to reflect the change of lipid content. The patients in this study have a long history of atherosclerosis, PVAT of carotid plaque has gone into the chronic phase, lipid fibrosis and microvascular remodeling occur, and the dynamic change of lipid content is small (11,30), so the ability of FAI to dynamically monitor vascular inflammation is limited (11). At the same time, Serum C reactive protein (CRP) is a marker of systemic inflammation and is associated with increased risk of stroke and unstable carotid atherosclerotic plaques (31). However, high sensitivity CRP is usually driven by other inflammatory conditions such as infection, arthritis, etc., and cannot specifically reflect the local inflammation of carotid atherosclerosis. PET is considered to be the most reliable non-invasive imaging modality for vascular inflammation. However, its clinical application is limited due to its low spatial resolution, high radiation exposure, and high cost. In our study, the RS of carotid PVAT was available, and the diagnostic efficacy of RS model of carotid PVAT in identifying symptomatic plaques was 0.745. RS analysis can accurately capture the texture changes of PVAT and reflect the level of vascular inflammation.The carotid plaque PVAT extracted in this study was extended 2cm above and below the center of the carotid segment bifurcation, with a total of 4cm as the longitudinal measurement distance.Because of the vascular shear stress (32), the vast majority of extracranial carotid plaques were distributed at the carotid bifurcation, and the plaques of the cases included in this study were distributed within 2cm above and below the carotid bifurcation. Second, there was little fat distribution around the carotid artery, and the plaques of the cases included in this study were distributed in the range of 2 cm above and below the carotid bifurcation. This study referred to the method of extracting the proximal 4cm PVAT of the coronary artery with pericoronary fat, and appropriately increased the collection range of the PVAT of carotid plaques to ensure the accuracy of RS extraction in PVAT. Thirdly, Oikonomou et al. showed (25) that perivascular FAI at 4cm proximal to the right coronary artery can reflect global coronary inflammation and predict cardiac mortality. In our study, PVAT at 4cm of the carotid bifurcation also has the potential to represent the risk of vascular inflammation at the carotid bifurcation plaque and the whole carotid artery segment.This paper has the following limitations: (1) This paper adopts the mainstream method used in current related research to identify symptomatic plaques, but it lacks a gold standard. In the future, we aim to collect plaque samples through carotid artery stripping and other procedures to accurately identify culprit plaques; (2) The lack of external validation datasets to evaluate the diagnostic efficacy of machine learning models; (3) As symptomatic and asymptomatic determination of plaques happens before CTA exams, it would have selection bias towards the model performance in the real clinical settings. In the next step of our research, we will conduct a prospective study on patients undergoing head and neck CTA to explore the association between PVAT imaging-based radiomics of carotid plaques and the occurrence of acute ischemic cerebrovascular events. The RS model of carotid plaque PVAT, when combined with the traditional feature model, demonstrates a significant improvement in the diagnostic performance for identifying symptomatic plaques compared to the traditional feature model alone. This indicates that the RS model of carotid plaque PVAT can serve as an objective indicator for evaluating plaque risk, providing a basis for risk stratification, as well as the diagnosis and treatment of carotid atherosclerotic diseases.The studies involving human participants were reviewed and approved by the Ethics Committee of Shunde Hospital, Guangzhou University of Chinese Medicine, approval number KY-2023072, and conforms to the ethical guidelines of the 1975 Declaration of Helsinki. The patients/participants provided their written informed consent to participate in this study.