105 Machine learning and carotid artery CT radiomics identify significant differences between culprit and non-culprit lesions in patients with stroke and transient ischaemic attack

• Published in: Heart (British Cardiac Society) Vol. 106; no. Suppl 2; pp. A82 - A84
• Main Authors: Le, Elizabeth Phuong Vi, Evans, Nicholas, Tarkin, Jason, Chowdhury, Mohammed, Zaccagna, Fulvio, Pavey, Holly, Wall, Chris, Huang, Yuan, Weir-McCall, Jonathan, Warburton, Elizabeth, Rundo, Leonardo, Schönlieb, Carola-Bibiane, Sala, Evis, Rudd, James HF
• Format: Journal Article
• Language: English
• Published: London BMJ Publishing Group LTD 01.07.2020
• Subjects: Accuracy; Carotid arteries; Confidence intervals; Flow velocity; Machine learning; Neural networks; Radiomics; Stroke; Transient ischemic attack
• DOI: 10.1136/heartjnl-2020-BCS.105
• ISSN: 1355-6037

IntroductionCarotid atherosclerosis is the main cause of ischaemic stroke. Texture analysis is a radiomic approach used to quantify image heterogeneity which can predict tumour aggression in oncology. We investigated whether this method could be applied to carotid artery disease to differentiate symptomatic from asymptomatic patients and culprit from non-culprit plaques, and then whether machine learning (ML) could correctly classify plaques based on these features.MethodsCT angiography (CTA) images from symptomatic patients with carotid artery-related cerebrovascular accidents (CVAs) and from asymptomatic (ASX) patients were studied. Regions-of-interest (ROIs) were drawn on 14 consecutive carotid artery CTA slices with 3mm slice thickness. PyRadiomics was used for isotropic image (1x1x1) resampling and normalisation prior to texture feature extraction from 6 different classes (Table 1). Asymptomatic carotids were compared to culprit carotids (CC), and non-culprit (NC) carotids using the Mann Whitney U test or Wilcoxon signed-rank tests as appropriate, with a p-value <0.0005 deemed statistically significant after Bonferroni correction. Non-normally distributed variables are reported as median (interquartile range). To assess the discriminatory ability of radiomic features in multi-class classification (ASX, CC or NC), texture features were fed into a Python scikit-learn pipeline for feature selection with variance thresholding, feature scaling and dimensionality reduction incorporated within a 5-fold cross validation scheme with 5 different ML classifiers to reduce the risk of data leakage and overfitting. Mean cross validation (CV) accuracy, area under the receiver operating curve (AUC) and 95% confidence intervals (CI) are reported.Abstract 105 Table 1Different classes of texture features extracted from carotid CTA Abbreviation Feature Class Number of Features IHFirst-order Intensity Histogram Statistics18GLCMGrey Level Co-occurrence Matrix24GLRLMGrey Level Run Length Matrix16GLSZMGrey Level Size Zone Matrix16GLDMGrey Level Dependence Matrix14NGTDMNeighbouring Grey Tone Difference Matrix5ResultsThe dataset comprised 82 carotid arteries from 41 symptomatic patients (41 culprit; 41 non-culprit) and 50 carotid arteries from 25 asymptomatic patients. CC and NC carotids showed significant differences in both first- and second-order features (IH Median: CC 618 (61); NC 646 (97), p<0.005) and (GLDM Large Dependence High Grey-Level Emphasis: CC 3147 (1837), NC 4811 (2181), p<0.0001), respectively. Both CC and NC carotids had higher heterogeneity than asymptomatic carotids (GLDM Dependence Entropy: CC 6.59 (0.43), NC 6.57 (0.52), ASX 6.24 (0.26), p<0.0001). All ML classifiers performed better than a randomly guessing classifier (mean accuracy 33.3%) in this multi-class (n=3) classification task (Table 2; Figure 2), with the neural network achieving the highest accuracy of 69%, CI [61%, 77%] with AUC 0.82 CI [0.78, 0.86].Abstract 105 Table 25-fold mean cross-validation auc with 95% confidence intervals Multi-Class ML Models AUC 95% confidence interval [CI] Decision Tree0.59[0.54, 0.64]Naïve Bayes0.75[0.73, 0.77]Random Forest0.79[0.70, 0.88]Logistic Regression0.80[0.77, 0.83]Neural Network0.82[0.78, 0.86]AUC, Area under the Receiver Operating Curve (‘OneVsRest’ scheme for multi-class classification tasks); ML, machine learning; CI, confidence intervalAbstract 105 Figure 1Rest and hyperaemic coronary flow velocity in the distal left anterior descending arteryAbstract 105 Figure 2Coronary flow reserve in controls and living kidney donors. Red squares represent mean. Error bars represent 95% confidence intervalsConclusionsTextural analysis combined with machine learning on carotid CT scans reveals highly significant differences between symptomatic and asymptomatic patients, and between culprit and non-culprit carotid arteries within symptomatic patients. This approach could help identify patients at high-risk of stroke for aggressive medical therapy and surveillance.Conflict of InterestNone