Table 1 Machine-learning-based SPECT dopaminergic imaging studies for PD diagnosis and early detection.
Study | Sample | Data features | Methods | Main findings | Other findings |
|---|---|---|---|---|---|
Acton and Newberg, 200615 | 81 PD, 94 HC | Striatum images of from [99mTc]TRODAT-1 SPECT images | Feature selection: Down-sample voxels in the striatum; Classification: ANN; Validation: Leave-one-out | Classification accuracy: 94.4% | ANN performed better than semi-quantitative ROI analysis (81.3%) and radiologists (88%); Difficult to interpret what ANN detect in the image |
Hamilton et al., 200670 | 18 PD (12 advanced PD, 6 early PD with ET) | Striatal uptake ratios (striatum-to-occipital cortex ratio and putamen-to-caudate tracer accumulation ratio) from 123I-FP-CIT SPECT images | Feature selection: none; Classification: ANN; Validation: Leave-one-out | Classification accuracy: 100% | Putamen-to-caudate tracer accumulation ratio is able to discriminate between PD and ET |
Palumbo et al., 201071 | 261 PD (89 ET, 64 early PD, 63 advanced PD) | Striatal uptake and uptake ratios (putamen/occipital, caudate/occipital) from (123)I-FP-CIT SPECT images | Feature selection: Down-sample voxels in the striatum; Classification: PNN and CIT; Validation: 50-fold-cross-validation | Classification accuracy: For PNN: Early PD: 81.9 ± 8.1%; Advanced PD: 78.9 ± 8.1%; ET: 96.6 ± 2.6% | CIT provided reliable cut-off values (e.g., 5.99 at putamen and 6.97 at caudate); Classification accuracy: For CIT: Early PD: 69.8 ± 5.3%; Advanced PD: 88.1%±8.8%; ET: 93.5 ± 3.4% |
Illan et al., 201244 | 108 PD, 100 HC | Striatal uptake image with normalized high intensity from 123I-ioflupane SPECT images | Feature selection: Applied a mask for high-intensity voxels; Classification: SVM, KNN, NM; Validation: Leave-one-third-out | Classification accuracy: AUC: 0.968 for SVM; 0.931 for KNN; 0.942 for NM | Classifier selection had higher impact on classification results than the preprocessing steps; Image preprocessing with voxel intensity normalized to a maximum value performed the best |
Segovia et al., 201245 | 95 PD, 94 HC | Striatal uptake with high intensity from 123I-ioflupane SPECT images | Feature selection: PLS and down-sampling; Classification: SVM; Validation: Leave-one-out | Classification accuracy: 94.7% (AUC: 0.968) | PLS + SVM outperformed previous approaches based on singular value decomposition |
Martinez-Murcia et al., 201447 | 158 PD, 111 HC | Computed Haralick texture features (via a gray-level co-occurrence matrix) from 123I-ioflupane SPECT images | Feature selection: none; Classification: SVM; Validation: Leave-one-out | Classification accuracy: 97.4% | |
Palumbo et al., 201446 | 56 PD, 34 non-PD | Uptake in the caudate (CL, CR) and putamen (PL, PR) from 123I-FP-CIT SPECT | Feature selection: none; Classification: SVM; Validation: Leave-one-out; fivefold cross-validation | Classification accuracy: CL + CR + PL + PR: 90.6% (leave-one-out); 90.7% (fivefold cross-validation) | Adding age improved classification accuracy (95.6%) |
Prashanth et al., 201416 | 369 early PD, 179 HC (from PPMI) | Striatal binding ratios from 123I-Ioflupane SPECT images | Feature selection: none; Classification: SVM; Validation: tenfold cross-validation | Classification accuracy: 96.14 ± 1.89% | SVM with non-linear kernel of radial basis function achieved higher classification accuracy than SVM with linear kernel (92%) |
Oliveira and Castelo-Branco, 201587 | 445 early PD, 209 HC (from PPMI) | Binding potential at each voxel in the striatum as feature extracted from 123I-ioflupane SPECT images | Feature selection: Apply BP threshold; Classification: SVM; Validation: Leave-one-out | Classification accuracy: 97.86% | The classification results were robust regardless the reference VOI used or the transformation used (for spatial normalization), or the way the features selected |
Hirschauer et al., 201537 | 189 PD, 62 SWEDD, 415 HC (from PPMI) | Motor and non-motor clinical features such as motor function and olfactory loss and imaging biomarkers such as ioflupane (123I) SPECT striatal-binding ratios | Feature selection: none; Classification: EPNN, PNN, SVM, KNN and CT; Validation: tenfold cross-validation | Classification accuracy: PD vs. HC: 98.6% (EPNN), 98.2% (KNN), 98.1% (SVM); PD vs. SWEDD: 95.3% (EPNN), 94.6% (KNN), 89.3% (SVM) | Classification accuracy: SWEDD vs. HC: 92% (EPNN), 91.6% (KNN), 89.3% (SVM) |
Huertas-Fernández et al., 201566 | 80 VP, 164 PD | Uptake in the striatum and whole brain from 123I-ioflupane SPECT images | Feature selection: t-test and Mann–Whitney test for the important features; Classification: LR, LDA, SVM; Validation: tenfold cross-validation | Classification accuracy: 90.3 ± 5.8% (LR for ROI approach); 90.4 ± 5.9% (SVM for voxel-based whole-brain approach) | Classification accuracy: 89.8 ± 6.5% (LDA for ROI), 89.9 ± 4.9% (SVM for ROI); 88.7 ± 4.9% (LR for voxel-based) 88.4 ± 6.4% (LDA for voxel-based) |
Prashanth et al., 201617 | 401 early PD, 183 HC (from PPMI) | Non-motor clinical features such as RBD and olfactory loss, CSF measurements and SPECT imaging markers (striatal-binding ratios) | Feature selection: none; Classification: SVM, random forests; Validation: tenfold cross-validation | Classification accuracy: 96.40 ± 1.08% for SVM, 96.18 ± 1.27% for random forests | SVM outperformed other classifiers; Combined features (non-motor clinical features and CSF and imaging markers) are useful for early detection of PD |
Choi et al., 201718 | 431 PD, 77 SWEDD, 193 HC (from PPMI); SNUH data: 72 PD, 10 HC | Striatal binding ratios and other imaging features from SPECT images | Feature selection: none; Classification: Deep CNN; Validation: tenfold cross-validation | Classification accuracy: 96% (PPMI); 98.8% (SNUH) | The performance of PD Net (deep CNN) was comparable to that of experts; SWEDD could be reclassified by PD Net |
Prashanth et al., 201788 | 427 early PD, 80 SWEDD, 208 HC (from PPMI) | Shape and surface-fitting-based features, striatal-binding ratios from SPECT images | Feature selection: Estimate feature importance with random forest; Classification: SVM, random forests; Validation: tenfold cross-validation | Classification accuracy (early PD vs. non-PD (SWEDD/HC)): 97.29 ± 0.11% for SVM, 96.9 ± 0.17% for random forests | SVM outperformed other classifiers; Shape and surface-fitting-based features showed higher importance than striatal-binding ratios for classification |
Wang et al., 201748 | 369 PD,165 NC (from PPMI). [93 AD, 202 MCI, 101 HC (from ADNI)] | PPMI: Striatal blinding ratios from SPECT images; Gray matter, white matter, and CSF volumes of ROIs from MRI images; ADNI: Gray matter volume of the ROIs from MRI images; mean intensity of ROIs from PET images | Feature selection: Optimization in progressive transductive learning; Classification: SVM, GTL; Validation: tenfold cross-validation | Classification accuracy: PPMI (PD vs. HC): SVM: 88.5% (MRI + SPECT); GTL: 97.4% (MRI + SPECT); ADNI (AD vs. HC): SVM: 86.7 ± 1.42% (MRI + PET); GTL: 92.6 ± 0.65% (MRI + PET) | Multi-modal features led to better classification performance than single-modal features |
Zhang and Kagen, 201749 | 1171 PD, 131 SWEDD, 211 HC (from PPMI) | A slice that has the highest striatal signal-to-background ratio of SPECT image was used | Feature selection: gradient descent optimization; Classification: Artificial Neural network; Validation: tenfold cross-validation | Classification accuracy: PD vs. HC: 93.8 ± 4.7% | A comparison of gradient descent and the Adagrad optimizer showed that there was no significant difference in their classification performance |
Taylor and Fenner, 201719 | 113 non-PDD, 191 PDD (Local data); 448 PD, 209 HC (from PPMI) | Voxel intensities; Principal components of image voxel intensities; Striatal binding radios (from the putamen and caudate) from (I123) Ioflupane (FP-CIT) SPECT images | Feature selection (data dimension reduction): PCA; Classification: SVM; Validation: tenfold cross-validation | Classification accuracy: semi-quantitative methods: 78~87% (local data), 89 ~95% (PPMI); SVM: 88~92% (local data), 95 ~97% (PPMI) | Machine-learning method performed better than semi-quantitative methods |
Taylor et al., 201820 | 304 PD (113 without PDD, 191 with PDD); 448 PD, 209 HC (from PPMI) | First five principal components of image voxel intensities in the striatum extracted from (123I)FP-CIT SPECT images | Feature selection: none; Classification: SVM; Validation: tenfold cross-validation | Classification accuracy: 92% | CADx increased the accuracy of the radiologists for research images, but had no significant change in accuracy for the clinical data and had less impact on the clinical scientist |
Oliveira et al., 201889 | 443 early PD, 209 HC (from PPMI) | Striatum uptake ratios and striatum dimensional-based features extracted from (123I)FP-CIT SPECT images | Feature selection: none; Classification: SVM; KNN; LR; Validation: Leave-one-out cross-validation | Classification accuracy: 97.9% (SVM, all features) | SVM outperformed other classifiers such as KNN and LR; Features with high classification accuracy: the length of the striatal region (96.5%), the putaminal binding potential (95.4%) and the striatal-binding potential (93.9%) |
