Post-revascularization Ejection Fraction Prediction for Patients Undergoing Percutaneous Coronary Intervention Based on Myocardial Perfusion SPECT Imaging Radiomics: a Preliminary Machine Learning Study

Abstract

In this study, the ability of radiomics features extracted from myocardial perfusion imaging with SPECT (MPI-SPECT) was investigated for the prediction of ejection fraction (EF) post-percutaneous coronary intervention (PCI) treatment. A total of 52 patients who had undergone pre-PCI MPI-SPECT were enrolled in this study. After normalization of the images, features were extracted from the left ventricle, initially automatically segmented by k-means and active contour methods, and finally edited and approved by an expert radiologist. More than 1700 2D and 3D radiomics features were extracted from each patient's scan. A cross-combination of three feature selections and seven classifier methods was implemented. Three classes of no or dis-improvement (class 1), improved EF from 0 to 5% (class 2), and improved EF over 5% (class 3) were predicted by using tenfold cross-validation. Lastly, the models were evaluated based on accuracy, AUC, sensitivity, specificity, precision, and F-score. Neighborhood component analysis (NCA) selected the most predictive feature signatures, including Gabor, first-order, and NGTDM features. Among the classifiers, the best performance was achieved by the fine KNN classifier, which yielded mean accuracy, AUC, sensitivity, specificity, precision, and F-score of 0.84, 0.83, 0.75, 0.87, 0.78, and 0.76, respectively, in 100 iterations of classification, within the 52 patients with 10-fold cross-validation. The MPI-SPECT-based radiomic features are well suited for predicting post-revascularization EF and therefore provide a helpful approach for deciding on the most appropriate treatment.

Keywords: Ejection fraction; Machine learning; Myocardial perfusion imaging; PCI; Quantitative features; Radiomics.

Fig. 1

An infographic flowchart summarizing different steps of the study from data retrieving to preprocessing, feature extraction, classification, and finally, performance evaluation of the proposed models

Fig. 2

Display of two of the features selected by NCA, where each column represents a patient from classes 1 to 3. These two features include Gabor and NGTDM. It visually depicts the image-based features of patients belonging to three classes

Fig. 3

NCA selected features’ correlation analysis. Smaller and brighter circles illustrate lower correlation values than larger and darker ones used for higher correlations. It indicates a low correlation for most of the features

Fig. 4

The seven selected features and their classes are depicted. Each row displays the feature normalized values for each patient, suggesting that values of features from specific classes tend to have similar values, while values of features from different classes tend to have roughly distinguishable values, highlighting the fact that classes are not properly classified

Fig. 5

Based on a combination of top features, consensus clustering is shown. This graph shows three grouped clusters without labels and clearly shows three distinct clusters of features

Fig. 6

Confusion matrix of the four best models depicting the mean of 100 times classification for each of the values

Fig. 7

Comparison of the performance of seven classifiers based on the NCA feature selection method using the Wilcoxon rank-sum test. The top panel illustrates the p-values obtained from comparing the evaluation metrics of seven classifiers, including accuracy, AUC, sensitivity, specificity, precision, and F-score. The lower panel displays the p values obtained from comparing the top four classifiers based on their evaluation metrics, compared on a per-class basis

Fig. 8

p Values obtained from the Wilcoxon rank-sum test comparing the performance of three feature selection methods, namely NCA, MRMR, and LASSO in seven different classifiers with evaluation metrics including accuracy, AUC, sensitivity, specificity, precision, and F-score

Fig. 9

ROC of cosine KNN, fine KNN, subspace KNN, and random forest classifiers for three classes. The AUC of the classifiers is greater than 0.5 for all three classes. In cosine KNN, the optimal point for class 1, class 2, and class 3, respectively, was 0.73, 0.83, and 0.58. Fine KNN’s optimal point was 0.78 for class 1, 0.82 for class 2, and 0.69 for class 3. In subspace KNN, the optimal point was 0.78 for class 1, 0.75 for class 2, and 0.81 for class 3. In random forest, the optimal point was 0.82 for class 1, 0.83 for class 2, and 0.50 for class 3. The x-axis represents the False Positive Rate (FPR) (1-specificity) and the y-axis represents the True Positive Rate (TPR) (sensitivity)