Identification of high-risk plaque features in intracranial atherosclerosis: initial experience using a radiomic approach

This study was approved by the Institutional Review Board of the Changhai Hospital (Shanghai, China), with all patients providing written informed consent. This study population was retrospectively selected from patients with intracranial stenosis who underwent intracranial HR-MRI between January 2014 and December 2016. All investigations were performed in accordance with the approved guidelines, including anonymisation of data during analysis. Patient inclusion criteria: (1) ischaemic stroke in basilar artery territory for the symptomatic group; (2) basilar artery stenosis on MRA or CTA; (3) ≥1 atherosclerotic risk-factors, including hypertension, diabetes mellitus, hypercholesterolaemia or cigarette smoking. Exclusion criteria: (1) non-atherosclerotic intracranial arterial disease, including vasculitis, moyamoya disease, dissection, reversible cerebral vasoconstriction syndrome and intracranial dolichoectasia; (2) chronic ischaemic stroke/transient ischaemic attack (TIA) (>12 weeks); (3) significant stenosis of the extracranial carotid arteries (> 30%) as assessed on ultrasound; (4) ascending aortic arch atheroma; (5) suspected cardio-embolic stroke; (6) known coagulopathy; (7) clinical contraindications to MRI, such as patients with pacemakers, certain metallic implants or severe claustrophobia.

Patients were scanned using a 3-T whole-body MRI scanner (Skyra; Siemens Healthcare, Erlangen, Germany) with a 20-channel phased array head and neck coil. Patients were scanned using two dimensional high-resolution black blood T1- and T2- weighted fast-spin-echo sequences. T1 images were acquired both pre- and post-contrast (gadolinium). After an initial multi-plane localiser sequence, axial three-dimensional (3D) time-of-flight (TOF) MR angiography was performed to identify the location of the basilar stenosis. Then HR-MRI scanning was performed in planes perpendicular to the longitudinal orientation of the artery. Scan parameters: both T1- and T2-weighted sequences were acquired with 12 × 2-mm-thick slices; field of view 100 mm × 100 mm; matrix 256 × 320; in-plane resolution of 0.4 mm × 0.3 mm. T1 images had TR/TE = 581 ms/18 ms, echo train length (ETL) = 4 and number of excitations (NEX) = 4. T2 images had TR/TE = 2,890 ms/46 ms; ETL = 20; NEX = 3.

All images were analysed by four radiologists (two radiology residents each with 3 years’ experience and two senior radiologists with 8 years’ experience in intracranial vessel wall imaging using these dedicated sequences). Each detected plaque was independently classified as acute/sub-acute symptomatic if a lesion was present on conventional neuroimaging (T2 FLAIR and DWI showed infarct) after an acute (<4 weeks) or sub-acute (4-12 weeks) ischaemic stroke/TIA with corresponding clinical symptoms. When multiple stenoses were present along the basilar artery, classification was based on the plaque at the location of maximal stenosis. A plaque was categorised as asymptomatic if detected in an asymptomatic patient and where conventional neuroimaging provided no evidence of infarct within the supplied vascular territory. Any disagreement was resolved by consensus.

Stenosis values were measured independently on HR-MRI images by one experienced radiologist (5 years’ experience in neuroradiology) [28]. Presence of fresh intraplaque haemorrhage (IPH) was identified if the signal intensity was >150% that of the nearby auricularis anterior muscles on pre-contrast T1-weighted images [29]. Lumen and outer wall boundaries were manually segmented on T2-weighted images at the slice with maximum plaque area using CMR Tools software (Cardiovascular Imaging Solutions, London, UK). Minimal luminal area (MLA) was computed based on the segmentation. MR signal intensity may vary within individuals due to factors such as coil positioning. Reproducibility of this area measurement method was previously reported to be excellent [30].

Gadolinium-based contrast agents were injected and post-contrast imaging was performed. The contrast enhancement ratio was measured at the slice of greatest enhancement, using adjacent grey matter (in a region of ~15 mm² at the hippocampus) to normalise the signal intensity. The contrast enhancement percentage was calculated as { (signal of plaque [post-contrast]/signal of grey matter [post-contrast] ) / (signal of plaque [pre-contrast]/signal of grey matter[pre-contrast]) – 1} × 100%. Plaque burden (PB) was measured at the site of maximal stenosis, and was defined as (1 – lumen area/outer area) ×100%.

The radiomic features were analysed by an experienced reviewer (5 years’ experience in vessel wall imaging) using dedicated software (3D-Slicer) [31]. To evaluate the reproducibility of the radiomic analysis, another reviewer (10 years’ experience in vessel wall imaging) independently measured the radiomic features in a subset of patients (n = 40) who were randomly selected from the study population. Each reviewer segmented the plaque boundaries at the level of the largest plaque area on T1, T2 and CE-T1 images. For each sequence, 94 radiomic features—including intensity (maximal intensity, mean intensity, standard deviation of intensity, etc), shape-based features (area, length, etc.), and textures—were extracted and analysed according to a previous publication [32]. Textures include grey-level co-occurrence matrix (GLCM), grey-level run-length matrix (GLRLM) and grey-level size-zone matrix (GLSZM). The textural features were described patterns or the spatial distribution of voxel intensities, which were calculated respectively from grey level co-occurrence (GLCM) and grey-level run-length (GLRLM) texture matrices. Determining texture matrix representations requires the voxel intensity values within the volume of interest (VOI) to be discretised. This discretisation step not only reduces image noise but also normalises intensities across all patients, allowing for a direct comparison of all calculated textural features between patients. The detailed definition of these features can be found in the previous publication [32]. The reproducibility of radiomic feature quantification was calculated by comparing the results from the two reviewers.

All statistical analyses were performed using MATLAB (R2013a; The Mathworks, Natick, MA, USA) and SPSS24.0. The mean and standard deviation (SD) were recorded for continuous variables, and frequency and percentage were recorded for categorical variables. For each variable, a normality test was performed. Univariate analysis was first performed to assess the association between each variable and acute symptomatic status. As all variables obtained from the radiomic analysis exhibited normal distributions, t-tests were used for the comparison of continuous variables. Chi-squared tests were used for the categorical variables. Multivariate analysis was then performed which included the variables that had p < 0.10 in the univariate tests. The odds ratios (ORs) with 95% confidence intervals (CIs) were calculated by a logistic regression model with stepwise selection of variables. Additionally, t-tests were used to first select the features with p < 0.05 in each sequence, and the features with significant differences and area under curve (AUC) values >0.65 were set as inputs for the random forest training features. To evaluate the diagnostic performance of radiomic features, supervised machine-learning methods using random forests [33] were applied to classify acute symptomatic from asymptomatic plaques.

A random forest is a meta-estimator that fits a number of decision-tree classifiers on various sub-samples of the dataset; ten classification trees with three layers were applied in this model; after each tree was voted the decision for classification result, the result was averaged to improve the predictive accuracy and control over-fitting. In random forests, each tree in the ensemble is built from a sample drawn with replacement from the training set. In addition, when splitting a node during the construction of the tree, the split that is picked is the best split among a random subset of the features.

The diagnostic performance was described using receiver operating characteristic (ROC) curves and AUC values. ROC curves were compared using the method developed by DeLong et al. [34] Reproducibility was evaluated by Bland-Altman plots and intraclass coefficient (ICC) using a two-way random model with absolute measurements. A p value <0.05 was considered statistically significant.

A total of 174 patients met the inclusion criteria. Seventy-eight patients were excluded due to intracranial aneurysm (n = 51), moyamoya disease (n =1 3), vasculitis (n = 7), dissection (n = 5) and bad image quality (n = 2). As a result, 96 patients were included in the final analysis. The demographic data of the patients is shown in Table 1. There were 43 acute, 18 sub-acute and 35 asymptomatic patients.

Clinical and radiological characteristics of the intracranial atherosclerotic plaques are summarised in Table 2. Sample patient images are shown in Figs. 1 and 2. Univariate analysis showed that gender, smoking, IPH, MLA and enhancement ratio were associated with acute/sub-acute symptomatic plaques (p < 0.1). Multivariate logistic regression analysis showed that IPH [odds ratio (OR) = 17.803; 95% CI, 2.093-151.472], MLA (OR = 1.515; 95% CI, 1.123-2.043) and enhancement ratio (OR = 71.979; 95% CI, 3.840-1349.211) were independent predictors of acute/sub-acute symptoms, and the AUC of the ROC curves was 0.638, 0.650 and 0.717, respectively (Fig. 3). The AUC value was improved to 0.833 when combining IPH, MLA and enhancement ratio in the model, with an optimised sensitivity of 84.8%, specificity of 71.0% and accuracy of 74.7%. Details of the diagnostic performance data are listed in Table 4.

Samples of the radiomic features are shown in Fig. 2. Random forest analysis showed seven radiomic features in T1 images and three features in CE-T1 images that were independently associated with acute/sub-acute symptoms (Table 3), while none of the features in T2 images were statistically significant. Diagnostic performance data of radiomics are listed in Table 4. T1 radiomic features had an AUC value of 0.893 and CE-T1 radiomic features, 0.918. When combining all the radiomic features from T1 and CE-T1 images, the AUC value improved to 0.936, with an optimised sensitivity of 97.0% and specificity of 79.0%, and accuracy of 83.2% (Fig. 3). Radiomic features had significantly higher AUC values than traditional features (p = 0.01).

Bland-Altman plots of traditional and radiomic measurements and the ICC values are shown in Supplementary Figs. 1 and 2. The ICCs of the traditional features for the two reviewers in measuring the degree of stenosis, PB, MLA and enhancement ratio were 0.782, 0.778, 0.875, 0.792 respectively (average ICC, 0.806). The Cohen’s kappa for the two reviewers in determining IPH was 0.811.

The average ICCs for the two reviewers in measuring the ten independent radiomic features were 0.836 (range, 0.798-0.881) for T1 images and 0.837 (range, 0.806-0.858) for CE-T1 images.

Although this study provided significant results, the sample size used to train the texture analysis of plaque was relatively small (n = 96). Future work would benefit from training with a larger data set. In addition, our analysis was performed using two-dimensional (2D) imaging data. It is possible that the use of 3D imaging acquisitions could better characterise plaque features for improved radiomic analysis. Also, reliance on a 2D TOF sequence to initially identify plaque may have underestimated the presence of non-stenotic atherosclerotic plaque. Three-dimensional imaging methods (up to 0.5 mm isotropic) [44] can reduce the partial volume effect of 2D acquisition and allow the detection of non-stenotic basilar artery or PCA plaques. However, the long scan time and wider point spread function induced by long echo trains must also be considered [45, 46]. Another potential improvement is the use of automated segmentation. Manual segmentation of the small intracranial plaque is still a challenging task. Automatic segmentation would reduce inter-observer variability and improve the feasibility of applying these methods to larger datasets.

This retrospective study was not able to ascertain the ability of radiomic analysis to predict the risk of future stroke risk, which is of paramount importance. Future prospective studies associating plaque features with long-term stroke risk through radiomic techniques could substantially shift the imaging paradigm from simple assessment of stenosis and subjective plaque components to a rigorous evaluation of objective and quantifiable plaque characteristics.

As a nascent field, radiomic methods may require additional refinement to improve accuracy. Diagnostic accuracy and predictive power are dependent on the quality of the data set utilised to train radiomic algorithms [23]. Leveraging quantitative imaging data with large population studies could provide actionable information to generate robust automated assessments of stroke risk based on vessel wall MRI of ICAD.

Radiomic analysis of basilar artery plaque texture on HR-MRI accurately distinguished between acute/sub-acute symptomatic and asymptomatic plaque. Radiomic features alone successfully predicted plaque symptom status over clinical imaging features in this retrospective study. Prospective studies are needed to further ascertain the ability of radiomic analysis of intracranial plaque HR-MRI to predict future stroke risk.