Coronary Artery Stenosis Detection using Myocardial Perfusion Imaging SPECT Radiomic Features and Machine Learning Algorithms

Introduction: Coronary stenosis is one of the leading causes of death due to cardiovascular disease in most countries worldwide. Therefore, early identification of evidence for this disease is essential to take the necessary measures to prevent further developments. This study aimed to diagnose coronary stenosis by radiomics analysis of myocardial perfusion imaging SPECT (MPI SPECT). For this purpose, the performance of multiple classifiers was evaluated to discriminate between normal and abnormal MPI SPECT in patients suspected of coronary artery stenosis.

Methods: A total of 183 patients acquired with MPI SPECT were selected. Image acquisition was performed for all patients with a 2-day stress-rest myocardial perfusion protocol. Only stress (exercise stress test, dipyridamole stress test) images were used in this study. Two nuclear medicine physicians with more than 5 years of experience reached a consensus to classify the images as normal and abnormal after the SPECT examinations (53 normal and 130 abnormal cases). Manual segmentation of volume-of-interest (VOI) for all images was performed using the 3D-slicer software package for whole cardiac. Then, radiomic features were extracted in 2D and 3D format under the guidelines of the Image Biomarker Standardization Initiative (IBSI) protocol using Standardized Environment for Radiomics Analysis (SERA) package. Random Forest Recursive Feature Elimination (RF-RFE) was used as feature selection algorithm. Performance of seven machine learning classifiers including gaussian Support Vector Machine (SVM), linear SVM, polynomial SVM, Logistic Regression (LR), Multi-Layer Perceptron (MLP), Naive Bayes (NB), and Random Forest (RF) were assessed. Data was randomly split to 70% and 30% for train and test sets while iterative resampling was used with replacement to create validation set during training procedure. Evaluation metrics including area under the ROC curve (AUC), accuracy, precision, recall, and f1_score were reported for test set.

Results: The morphological and distance zone matrix features (GLDZM) had the highest contribution to the prediction of patients with coronary stenosis. Among the classifiers, the random forest model provided the best performance (AUC: 0.73, accuracy: 0.78, precision: 0.81, recall: 0.93, and f1_score: 0.86).

Conclusions: In this work, we showed that employing radiomics modeling on myocardial perfusion imaging SPECT can be used as decision-support by nuclear medicine physicians to accurately discriminate between normal and abnormal cases of coronary stenosis. It can reduce the labor and analysis time in nuclear medicine clinics. Besides, random forest algorithm showed superior performance in comparison with other classifiers.

• Download figure
• Open in new tab
• Download powerpoint

• Download figure
• Open in new tab
• Download powerpoint