Skip to main content

Table 8 Summary of the available literature on machine learning algorithms for binary classification of (I123)FP-CIT images since 2010, ordered according to maximum accuracy (where available)

From: Comparison of machine learning and semi-quantification algorithms for (I123)FP-CIT classification: the beginning of the end for semi-quantification?

Authors Image features Classifier Validation data + method Results
Augimeri et al. 2016 [30] Mean ellipsoid uptake, dysmorphic index (ellipsoid orientation) SVM 43 local images (12 normal, 31 Parkinson’s disease (PD)), no cross-validation mentioned Up to 100% accuracy, specificity and sensitivity
Bhalchandra et al. 2015 [31] Analysis of 42nd slice only. Striatal binding ratios in both caudates and putamina, radial features and gradient features. Features are tested for statistical significance (wilcoxon rank) before use in the classifier Linear SVM and SVM with Radial Basis Function (RBF) kernel, Linear Discriminant Analysis (LDA) 350 images from PPMI database (187 healthy controls (HC), 163 PD). 5 fold cross-validation (CV), repeated 100 times Linear SVM: maximum of accuracy = 99.4%
RBF kernel: maximum of accuracy = 99.4%
LDA: maximum of accuracy = 99.4%
Oliveira and Castelo-Branco 2015 [32] Image voxels within striatal region of interest Linear SVM 654 images from PPMI database (209 HC, 445 PD). Leave-one-out CV Maximum of accuracy = 97.9%
Sensitivity = 97.8%
Specificity = 98.1%
Prashanth et al. 2017 [33] 16 shape and 14 surface fitting features of selected slices, following thresholding. Striatal binding ratios of both caudates and putamina and asymmetry indices were also considered. Features are tested for statistical significance (wilcoxon rank) before use in the classifier SVM with RBF kernel, boosted trees, random forests, naive bayes 715 images from PPMI database (208 HC, 427 PD, 80 scans without evidence of dopaminergic deficit (SWEDD)). 10 fold CV, repeated 100 times. Hyperparameters for SVM chosen through 10 fold CV SVM: accuracy = 97.3 ± 0.1%
Sensitivity = 97.4 ± 0.1%
Specificity = 97.2 ± 0.2%
Boosted trees: accuracy = 96.8 ± 0.2%
Sensitivity = 97.1 ± 0.3%
Specificity = 96.3 ± 0.4%
Random forests: accuracy = 96.9 ± 0.2%
Sensitivity = 97.2 ± 0.2%
Specificity = 96.5 ± 0.3%
Naive Bayes: accuracy = 96.9 ± 0.1%
Sensitivity = 96.4 ± 0.1%
Specificity = 96.5 ± 0.2%
Tagare et al. 2017 [34] Voxel intensities within a region of interest Logistic lasso 658 images from PPMI database (210 HC, 448 PD). 3 fold CV for performance assessment. Parameters chosen through 10 fold CV (nested within outer 3 fold CV). Maximum of accuracy = 96.5 ± 1.3%
Palumbo et al. 2014 [35] Striatal binding ratios for both caudates and putamina (and a subset of these 4 features), patient age SVM with RBF kernel 90 local images from patients with ‘mild’ symptoms (34 non-PD, 56 PD). Leave-one-out and 5 fold CV Maximum of accuracy = 96.4%
Prashanth et al. 2014 [36] Striatal binding ratio for both caudates and putamina SVM, linear and RBF kernel. 493 images from PPMI database (181 HC, 369 early PD), 10 fold CV, no repeats RBF kernel: accuracy = 96.1%, sensitivity = 96.6%, specificity = 95.0%
Linear SVM: accuracy = 92.3%, sensitivity = 95.3%, specificity = 84.0%
Martinez-Murcia et al. 2013 [37] 12 Haralick texture features within a brain region of interest Linear SVM ‘Whole’ PPMI database. Leave-one-out CV Maximum of accuracy = 95.9%, sensitivity = 97.3%, specificity = 94.9%
Zhang and Kagen 2016 [38] Voxel intensities from a single axial slice, repeated for 3 different slices Single layer Neural network 1513 images from PPMI database (baseline and follow-up, 1171 PD, 211 HC, 131 SWEDD). 1189 images for training, 108 for validation, 216 for testing. 10 fold CV Maximum of accuracy = 95.6 ± 1.5%, sensitivity = 97.4 ± 4.3%, specificity = 93.1 ± 3.6%
Rojas et al. 2013 [39] Voxel intensities, independent component analysis (ICA) & principal component analysis (PCA) decomposition of voxel data (after applying empirical mode decomposition) within regions of interest Linear SVM 80 local images (39 non-pre-synaptic dopaminergic deficit (non-PDD), 41 PDD). Leave-one-out CV Raw voxels: accuracy = 87.5%, sensitivity = 90.2%, specificity = 84.6%
ICA features: maximum of accuracy = 91.2%, sensitivity = 91.8%, specificity = 92.9%
PCA features: maximum of accuracy = 95.0%, sensitivity = 95.1%, specificity = 94.9%
Towey et al. 2011 [12] PCA decomposition of voxels within striatal region of interest Naïve-Bayes, Group prototype 116 local images (37 non-PDD, 79 PDD). Leave-one-out CV Naïve-Bayes: accuracy = 94.8%, sensitivity = 93.7%, specificity = 97.3%
Group prototype: accuracy = 94.0%, sensitivity = 93.7%, specificity = 94.6%
Segovia et al. 2012 [40] Partial least squares decomposition of voxels within striatal regions SVM applied to hemispheres separately. RBF kernel 189 local images (94 non-PDD, 95 PDD). Leave-one-out CV Features varied from 1 to 20. Maximum of accuracy = 94.7%, sensitivity = 93.2%, specificity = 93.6%
Martinez-Murcia et al. 2014 [41] ICA decomposition of selected voxels SVM, linear and RBF kernel 208 local images (100 non-PDD, 108 PDD), 289 images from PPMI database (114 normal, 175 PD). 30 fold CV RBF kernel: maximum of accuracy = 94.7%, sensitivity = 98.1%, specificity = 92.0%
Linear SVM: maximum of accuracy = 92.8%, sensitivity = 98.2%, specificity = 93.0%
Illan et al. 2012 [42] Image voxel intensities and image voxels within striatal region of interest Nearest mean, k-nearest neighbour (k-NN), linear SVM 208 local images (108 non-PDD, 108 PDD). 30 random permutations CV, with 1/3 data held out for testing SVM: maximum of sensitivity = 89.0%, specificity = 93.2%
Nearest mean: maximum of sensitivity = 90.7%, specificity = 84.0%
k-NN: maximum of sensitivity = 88.6%, specificity = 86.9%
Palumbo et al. 2010 [43] Striatal binding ratios for caudate and putamina on 3 slices Probablistic neural network (PNN), Classification tree (CT) 216 local images (89 non-PDD, 127 PD). Two fold CV, repeated 1000 times PNN: for patients with essential tremor mean probability of correct classification = 96.6 ± 2.6%
CT: for patients with essential tremor mean probability of correct classification = 93.5 ± 3.4%
  1. Algorithms using only (I123)FP-CIT SPECT data are considered, multimodal inputs are excluded. Literature lacking accuracy data are grouped at the bottom of the table