Coronary artery calcification may lead to significant stenosis of the coronary artery and is hence able to decrease oxygen supply to the heart. Coronary sclerosis may have a chronic effect on myocardial perfusion, and therefore changes of the myocardium may be conceivable depending on the level of coronary calcification. It is assumed that novel Photon Counting CT may improve the detection of textural changes due to improved spatial resolution and lower contrast-to-noise ratio. Radiomics analysis lends itself as an analysis technique as it converts a region of interest (ROI) into mineable high-dimensional data and thus derives information not visible to the human eye from the radiological image that may aid in automatic classification.

 

In this work, the effect of the severity of coronary artery calcification, measured by the Agatston score, on left ventricular myocardial texture features is determined. In order to achieve that, a complete radiomics pipeline including a classifier was implemented to classify Agatston scores based on myocardial radiomics features.

The Agatston score was specified for 30 patients who underwent PCCT scans between December 2021 and February 2022. The Agatston score was divided into three classes, with 10 representatives each. Class A0 contains the Agatston scores of zero, A1 holds the scores of 1 to 99, while A2 contains values greater than 100. The cut-off value of 100 was chosen because calcium levels above this value can be associated with most coronary events due to coronary artery disease. The entire left ventricular myocardium, including the trabecular structure and papillary muscle, was defined as the ROI and segmented semi-automatically.

After gray level normalization and radiomics feature extraction, different feature selection methods and classifiers were tested. Feature scaling was applied using the z-score. For the classifiers, a Leave-One-Out cross-validation and hyperparameter tuning was performed.

 

An Extra-Trees classifier for feature selection achieved the highest results in combination with a Support Vector Machine. Patients' Agatston scores could be correctly predicted using 4 radiomics features with an overall accuracy of 0.73 ± 0.158 at the 95% confidence interval. Even though all classes exhibit similar results, small differences can be seen. Class A0 has the highest precision because it has the lowest false positive rate 2/10. A1 has the highest recall, because it has the fewest false negatives 2/10 and the highest true positives 8/10. F1-Score is highest for this class as well. The error metrics of A2 are well balanced with the same precision, recall, and F1-Score because the false positives and false negatives are both three.

 

 

 

Figure 1: The left clustermap shows all of the 107 extracted features. The right clustermap shows the four remaining features as the result of the feature selection and the Agatston score.

 

 

Figure 2: The classification report of the SVM. The results do not show much difference between the classes.

 

This preliminary study indicates that the effect of coronary artery calcification on the texture analysis of the left ventricular myocardium, indicating a possible structural change of the myocardium as a function of the Agatston score, could be identified.