Comparison of 12 Machine Learning Methods for Polar Map Classification in Cardiac Perfusion PET

We evaluated 12 open source machine learning methods for classification of polar map images in cardiac perfusion positron emission tomography (PET) using a dataset consisting of 138 polar maps. Majority of the classifiers showed good accuracy in 10-fold cross-validation (mean accuracy of 0.75–0.88). Accuracy was slightly lower when applied to a separate hold-out dataset (0.70-0.87). From the evaluated classifiers, a support vector machine using a polynomial kernel, Gaussian naive Bayes classifier and a two-dimensional convolutional neural network had stable performance across both the cross-validation and hold-out datasets (accuracy of 0.78, 0.83 and 0.87).