Automated Classification of Atherosclerosis in Coronary Computed Tomography Angiography Images Based on Radiomics Study Using Automatic Machine Learning

Coronary computed tomography angiography (CCTA) has been recognized as a widely used non-invasive coronary imaging approach, which provides the assessment of luminal stenosis. However, the current interpretation of CCTA images still depends on qualitative assessment, which is prone to subjective variability and considered time-consuming. Multiple studies were conducted, venturing into the application of machine learning models specifically used for the classification of atherosclerotic plaques. Hence, this experimental study was designed to classify the atherosclerotic plaques from CCTA images using Auto-WEKA. In this study, there were 202 patients’ original CCTA images collected retrospectively from Institut Jantung Negara (IJN). Semi-auto segmentation of three main coronary arteries was performed on the axial view of CCTA multi-slice images which resulted in a sum of 606 Volume of Interest (VOI). The radiomic features included the first-order, second-order, and shape-order features were extracted from each VOI and acted as an input dataset for the automated machine learning (AutoML) tool which was Auto-WEKA to perform the classification as either normal, calcified, mixed, or non-calcified atherosclerotic plaques. In this study, the best classifier suggested among 39 machine learning methods tested by Auto-WEKA was the random forest. The classification performance was evaluated in terms of multi-class classification of confusion matrix, recall (sensitivity), precision (PPV), F-measure, inverse F-measure, accuracy, and receiver operating characteristics (ROC) curve as well as area under the curve (AUC). Overall, the results showed the highest accuracy of 87% (F-measure: 0.69; Inverse F-Measure: 0.92; AUC: 0.9278) in classifying the calcified plaques using the best classifiers suggested by Auto-WEKA compared to normal, non-calcified and mixed plaques. For the normal plaques, it showed the accuracy of 83% (F-measure: 0.85; Inverse F-Measure: 0.80; AUC: 0.9172), while the non-calcified and mixed plaques showed the accuracy of 77% (F-measure: 0.43; Inverse F-Measure: 0.85; AUC: 0.7911) and 80% (F-measure: 0.54; Inverse F-Measure: 0.87; AUC: 0.7986), respectively. In conclusion, Auto-WEKA showed promising results in obtaining the best classifier among 39 machine learning for the classification of the calcified plaques compared to normal, non-calcified, and mixed plaques based on a CCTA-based radiomic dataset.