Table 3.
Machine-learning-based structural MRI studies for PD diagnosis and early detection.
| Study | Sample | Data | Methods | Main findings | Other findings |
|---|---|---|---|---|---|
| Duchesne et al., 200975 | 16 PD, 8 probable PPS, 8 probable MSA, 149 HC | Intensity and shape-based features for brain tissue composition and deformation in the hindbrain region from MRI | Feature selection: PCA; Classification: SVM with least-squares optimization; Validation: leave-one-out | Classification accuracy (PD vs. PSP or MSA): 91% (sensitivity 79~87%, specificity 87~96%) | Automatic imaging feature extraction and classification may aid in the diagnosis of PD vs. PSP or MSA |
| Focke et al., 2011b76 | 21 PD, 10 PSP, 11 MSA, 22 HC | GM and WM volume from MRI (by VBM) | Feature selection: threshold images; Classification: SVM; Validation: leave-one-out | Classification accuracy: (PD vs. PSP) 87.1% for GM 96.8% for WM; (PD vs. MSA) 71.9% for GM 65.63% for WM | GM and WM volume did not differentiate PD from HC |
| Haller et al., 201284 | 17 PD, 23 other Parkinsonism (5 MSA; 1 PSP; 17 other types) | TBSS from DTI | Feature selection: select the most discriminative features with RELIEF; Classification: SVM; Validation: tenfold cross-validation | Classification accuracy: 97.5 ± 7.54% (PD vs. other Parkinsonism) | PD had a spatially consistent increase in FA and decrease in MD in the right frontal white matter |
| Haller et al., 201378 | 16 PD, 20 other Parkinsonism | SWI | Feature selection: select the most discriminative features with RELIEF; Classification: SVM; Validation: tenfold cross-validation | PD had increased SWI in the bilateral thalamus and left substantia nigra; Classification accuracy: 86.92 ± 16.59% (PD vs. Other) | Visual analysis yielded no differences between groups |
| Salvatore et al., 201477 | 28 PD, 28 PSP, 28 HC | Imaging features obtained by PCA; Voxel-based pattern distribution map of structural differences from MRI | Feature selection: PCA; Classification: SVM; Validation: leave-one-out | Classification accuracy (Specificity/Sensitivity): 93.5 (90.6/97.4)% for PD vs HC; 92.2 (92.5/92.4)% for PSP vs HC; 92.2 (91.3/94.4)% for PSP vs PD | Regions in the midbrain, pons, corpus callosum and thalamus |
| Cherubini et al., 201485 | 57 probable PD, 21 PSP (9 with probable PSP and 12 with possible PSP) | GM and WM volumes from MRI; FA and MD from DTI; DAT-SPECT used as ground truth | Feature selection: F-test for the important ROIs; Classification: SVM; Validation: Leave-one-out | Classification accuracy: All features combined: 100%; GM + MD + FA: Sensitivity: 90%; Specificity: 96% | Classification accuracy: Sensitivity: 76% (GM), 100% (WM), 86% (FA), 57% (MD); Specificity: 93% (GM), 100% (WM), 88% (FA), 93% (MD) |
| Singh and Samavedham, 201526 | 518 early PD, 68 SWEDD, 245 HC (from PPMI) | Voxel intensity change images, and GM and WM volumes of 500 ROIs from MRI (by KSOM) | Feature selection: WAT; Classification: Least-squares SVM; Validation: 20-fold cross-validation | Classification accuracy: PD vs. HC: 93.25 ± 0.46% for GM; 96.84 ± 0.28% for WM; PD vs. SWEDD: 99.86 ± 0.1% for GM 98.59 ± 0.48% for WM; SWEDD vs. HC: 100 ± 0% for GM 99.21 ± 0.36% for WM | Compared with HC, PD had atrophy in regions such as putamen, thalamus, and corpus callosum; Volume loss in regions such as cerebellum may help differentiate e SWEDD with PD |
| Dinov et al., 201636 | 263 PD, 40 SWEDD, 127 HC (from PPMI) | Clinical data (e.g., UPDRS scores), demographic data (e.g., age), genetics data (e.g., chr12), and neuroimaging biomarker (e.g., cerebellum shape index) from MRI | Feature selection: hillclimbing search, CARET; Classification: Model-based such as GLM and MMRM; Data-driven: AdaBoost, SVM, Naïve Bayes, Decision Tree, KNN, K-Means; Validation: fivefold cross-validation | Classification accuracy: PD vs. HC: 96.2% for SVM, 98.9% for AdaBoost; PD + SWEDD vs. HC: 94.5% for SVM, 98.3% for AdaBoost | Model-free or data-driven methods outperformed model-based methods; Including UPDRS data improved classification accuracy |
| Huppertz et al., 201674 | 204 PD, 73 HC, 106 PSP, 21 MSA | Volumetric measures (of 44 ROIs in GM, WM, CSF, brain lobes, cerebellum, midbrain, etc.) from MRI | Feature selection: none; Classification: SVM; Validation: Leave-one-out | Atrophy in the midbrain, basal ganglia, and cerebellar peduncles contributed most to classification; Classification accuracy: PD vs. HC: 66.2%; PSP vs. HC: 91.4%; MSA vs. HC: 82.4-88.4%; Multi-class classification: PD: 86.9%; PSP: 85.4%; MSA: 87.2% | Midbrain atrophy is the hallmark of PSP; Atrophy in pons is most prominent in MSA; PD had subtle volume reduction in cerebral GM (esp. basal ganglia) |
| Adeli et al., 201690 | 374 PD, 169 HC (from PPMI) | GM, WM volumes of 98 ROIs from MRI | Feature selection: JFSS; Classification: Robust LDA; SVM; Validation: Leave-one-out | Classification accuracy: 81.9% for Robust LDA; 69.1% for SVM | Feature selection with JFSS and classification with robust LDA outperformed other feature selection and classification methods; This approach can be applied to other neurodegenerative disorders |
| Gu et al., 201680 | 52 PD (19 PIGD, 25 TD, 8 mixed subtype), 45 HC | GM, WM, CSF volumes from MRI; FA, MD, RD, AD from DTI; ReHo and ALFF from resting-state fMRI | Feature selection: Recursive feature elimination; Classification: SVM; Validation: Leave-one-out | Classification accuracy: PIGD vs. non-PIGD 92.3% | The diagnostic agreement evaluated by the Kappa test showed Kappa value = 0.83 for agreement with the existing clinical categorization |
| Zeng et al., 201752 | 45 probable PD, 40 HC | GM in the cerebellum from MRI | Feature selection: Recursive feature elimination; Classification: SVM; Validation: Leave-one-out; fivefold (twofold, 632-fold) cross-validation | Classification accuracy: 97.7% for leave-one-out validation, 97.2% and 96.9% for twofold and fivefold cross-validation respectively | PD had GM density decrease in the Crus and Vermis of the cerebellum |
| Du et al., 201786 | 35 PD, 36 HC, 16 MSA, 19 PSP | DTI (FA, MD) and the R2* (apparent transverse relaxation rate) measures in the striatal, midbrain, limbic, and cerebellum | Feature selection: Regularized logistic regression; Classification: Elastic-Net machine learning and receiver-operating characteristic curve analysis; Validation: nested tenfold cross-validation | Classification accuracy: PD vs. HC: 91% (DTI + R2*), 82% (DTI); PD vs. MSA: 99% (DTI + R2*), 89% (DTI); PD vs. PSP: 99% (DTI + R2*), 97% (DTI); MSA vs. PSP: 98% (DTI + R2*), 96% (DTI); | MSA showed decreased FA and an increased R2* in the subthalamic nucleus, whereas PSP showed an increased MD in the hippocampus |
| Peng et al., 201751 | 69 PD, 103 HC (from PPMI) | GM, WM, CSF volumes, cortical thickness, cortical surface area, correlation index of cortical thickness of 78 ROIs | Feature selection: Recursive feature elimination; Classification: SVM; Validation: tenfold cross-validation | Classification accuracy: 85.8% (combined all features), 71.6% (GM + WM + CSF) | The most sensitive features are in the frontal lobe, parental lobe, limbic lobe, temporal lobe, and central region |
| Amoroso et al., 201825 | 374 PD 169 HC (from PPMI) | Network measures (correlation of patch voxel intensity distribution) from MRI images, and clinical data | Feature selection: Random forest; Classification: SVM; Validation: tenfold cross-validation | Classification accuracy (AUC): 0.88 ± 0.06 (MRI network measures); 0.70 ± 0.08 (clinical data); 0.93 ± 0.04 (combined features) | This MRI network approach has better classification accuracy than VBM (0.86 ± 0.06) and ROI (0.72 ± 0.07). |
| Singh et al., 201827 | 408 PD 71 SWEDD (from PPMI); [128 AD 262 HC 447 MCI (from ADNI)] | Discretized Voxel Intensity Changes extracted from MRI images by SOM | Feature selection: SOM; Classification: SVM; Validation: tenfold cross-validation | Classification accuracy: 92.63 ± 0.06% (PD vs. HC); 94.63 ± 0.05% (PD vs. SWEDD); 92.65 ± 0.08% (SWEDD vs. HC); [94.29 ± 0.08% (AD vs. HC); 85.43 ± 0.08% (AD vs. MCI); 95.24 ± 0.05% (MCI vs. HC);] | Biomarkers were identified to further identify clinically relevant ROIs for differential diagnosis |
ADNI Alzheimer’s disease neuroimaging initiative, AUC area under the receiver-operating characteristic (ROC) curve, PD Parkinson disease, GM gray matter, WM white matter, CSF cerebrospinal fluid, FA fractional anisotropy, MD mean diffusivity, RD radial diffusivity, AD axial diffusivity, PSP progressive supranuclear palsy, MCI mild cognitive impairment, MSA multisystem atrophy, TBSS tract-based spatial statistics, VBM voxel-based morphometry, PCA principal components analysis, PPMI Parkinson’s progression markers initiative, SVM support vector machine, SWEDD scans without evidence of dopaminergic deficit, KSOM Kohonen self-organizing map, WAT Welch–Aspin test, UPDRS unified Parkinson’s Disease Rating Scale, GLM generalized linear models, MMRM mixed-effect modeling with repeated measurements, GEE generalized estimating equations, JFSS joint-feature sample selection, ReHo regional homogeneity, ALFF amplitude of low-frequency fluctuation, PIGD postural instability and gait difficulty, RBD rapid eye movement (REM) sleep behavior disorder, ROI region of interest, TD tremor-dominant, SOM self-organizing maps.