Fig. 1

Feature selection, parameter selection, classification, and reproducibility framework. To find the best classification model, the framework is started with single photon emission computerized tomography (SPECT) scans (162 subjects) (a). Information Gain algorithm (b) removes non-informative voxels. A loop of parameter selection and Support Vector Machines (SVMs)-based feature selection then takes place. Only voxel clusters with size ≥20 are kept in the dataset (c) with DBSCAN, a density-based clustering method. At each iteration, the dataset is trained and tested (d); SVM’s feature elimination (e) refines voxels before next DBSCAN run. When there was no more than 100 voxels, parameter search was ended. At the next steps, model, 10-fold Cross Validation (10xCV), and leave-one-out (LOO) classifications were carried out and accuracies and set of selected voxels were identified (f)