INTRODUCTION

Intracranial atherosclerotic disease (ICAD) is one of the main causes of ischemic stroke, which is the leading cause of death and disability worldwide ^[1]. As a systemic disease, risk factors for symptomatic and asymptomatic ICAD are complex. Increasing evidence supports that vulnerable plaque is correlated with risk for stroke, which reveals the importance of intracranial plaque risk identification. Magnetic resonance imaging (MRI) has been established as a reliable and non-invasive clinical method for plaque characterization. Due to the MRI can provide a variety of vessel wall information, experts need to evaluate quantitatively and qualitatively. Despite of these visual information, some additional features based on data statistics are also benefit for building risk prediction model. Radiomics is an automated and repeatable approach for extracting massive features for medical imaging. However, few articles have focused on radiomic-based studies of intracranial plaque of basilar artery and middle cerebral artery. In this study, we propose to build a high-risk intracranial plaque model using radiomics features and machine learning to differentiate symptomatic plaque from asymptomatic plaque, which is helpful in guiding clinical management.

METHODS

• Patient inclusion criteria

Study population was recruited from Xuanwu Hospital (Beijing, China) who suffered from symptomatic and asymptomatic ICAD and underwent whole-brain vessel wall MRI. Study participants retrospectively selected from patients with intracranial stenosis: at least one ≥50% intracranial artery as confirmed by MR angiography, computed tomography angiography, or digital subtraction angiography. According to whether they had clinical symptoms, we divided the participants into two groups. The clinical symptoms are defined as: ischemic stroke in the territory of stenotic intracranial artery within the past 30 days. otherwise, no history of stroke in the territory of stenotic intracranial artery is defined as asymptomatic. The exclusion criteria included: (1) coexistent ipsilateral internal carotid stenosis; (2) preexistin conditions such as vasculitis, moyamoya disease, dissection, reversible cerebral vasoconstriction syndrome; (3) evidence of cardioembolism. A total of 91 patients were included, therein 60 symptomatic and 31 asymptomatic.

• MRI acquisition and segmentation

Whole-brain vessel wall MRI was performed at 3T Siemens Verio. The image protocol settings and patient demographic data are shown in Table I. The plaque segmentation on T1 and CE-T1 images was performed by an experienced reviewer on an open-source software ITK-SNAP (version 3.8.0, www.itk-snap.org). Figure 1 demonstrates the segmentation plaque samples. Figure 1(a) presents the Maximum Intensity Projection(MIP), precontrast T1 and postcontrast T1 sequences of middle cerebral artery from a symptomatic patient. The basilar artery sample from asymptomatic group is shown in Figure 1(b).

• Model development

Before the statistical analysis, we do the full brain normalization for each contrast so that the intensity of imaging is rescaled to 0-255. In our study, 2264 radiomic features were obtained on each contrast image by Pyradiomics platform^[2] via five image filters and six feature classes. We perform z distribution normalization to radiomic features before dimension reduction. For each feature, levene test and t-test are performed to assess the association between symptomatic group and asymptomatic group. Selected values with p<0.05 are set as inputs for least absolute shrinkage and selection operator (LASSO) algorithm, which aims for further feature reduction. Finally, totally 26 features are included for machine learning method training. The correlation heat map among these features is demonstrate in Figure 2. Five basic machine learning models: support vector machine (SVM), random forest (RF), linear discriminant analysis (LDA), Ridge, multi-layer perceptron (MLP) are developed to classify symptomatic from asymptomatic plaques. The variates from precontrast T1(n=12) and postcontrast T1(n=14) are respectively introduced to the machine learning method and generate the risk probabilities, which respectively are defined as p_pre and p_pre. are calculated as p=p_pre*λ+p_post*(1-λ). In this study, we set λ as 0.7. Figure 3 show the flowchart of study process.

RESULTS

We use a 5-fold cross validation to evaluate the performance of the machine learning models and also calculate the average receiver operating characteristic (ROC) analysis and the area under the ROC curve (AUC). The cross-validation results and the mean ROC curve are shown in Figure 4. As shown in results, when combining all the radiomic features from precontrast T1 and postcontrast T1 images, the AUC value improved. We also test all features on multiple machine learning methods. The SVM method shows the best performance with mean AUC of 0.88 compared with other familiar regression models.

DISCUSSION and CONCLUSION

In this study, we adopt radiomics to acquire richer information and focus on high-risk intracranial plaque, which is rarely investigated by radiomics analysis. The model consists of 26 quantitative features after dimension reduction and achieves excellent diagnostic performance, which are not perceived by radiologists. However, the relatively small number of patients and single medical center are still inconvenienced the performance and stability of method. Our future work will focus on combining model of traditional qualitative indicators, radiomics features and deep learning based information from images.

Acknowledgements

No acknowledgement found.