Machine-learning integration of CT histogram analysis to evaluate the composition of atherosclerotic plaques:...

To determine whether machine learning with histogram analysis of coronary CT angiography (CCTA) yields higher diagnostic performance for coronary plaque characterization than the conventional cut-off method using the median CT number.

We included 78 patients with 78 coronary plaques who had undergone CCTA and integrated backscatter intravascular ultrasound (IB-IVUS) studies. IB-IVUS diagnosed 32 as fibrous- and 46 as fatty or fibro-fatty plaques. We recorded the coronary CT number and 7 histogram parameters (minimum and mean value, standard deviation (SD), maximum value, skewness, kurtosis, and entropy) of the plaque CT number. We also evaluated the importance of each feature using the Gini index which rates the importance of individual features. For calculations we used XGBoost. Using 5-fold cross validation of the plaque CT number, the area under the receiver operating characteristic curve of the machine learning- (extreme gradient boosting) and the conventional cut-off method was compared.

The median CT number was 56.38 Hounsfield units (HU, 8.00-95.90) for fibrous- and 1.15 HU (-35.8-113.30) for fatty- or fibro-fatty plaques. The calculated optimal threshold for the plaque CT number was 36.1 ± 2.8 HU. The highest Gini index was the coronary CT number (0.19) followed by the minimum value (0.17), kurtosis (0.17), entropy (0.14), skewness (0.11), the mean value (0.11), the standard deviation (0.06), and the maximum value (0.05), and energy (0.00). By validation analysis, the machine learning-yielded a significantly higher area under the curve than the conventional method (area under the curve 0.92 and 95%, confidence interval 0.86-0.92 vs 0.83 and 0.75-0.92, p = 0.001).

The machine learning-was superior the conventional cut-off method for coronary plaque characterization using the plaque CT number on CCTA images.