Table 2.
Machine-learning-based PET imaging studies for PD diagnosis and early detection.
| Study | Sample | Data features | Methods | Main findings | Other findings |
|---|---|---|---|---|---|
| Segovia et al., 201567 | 39 PD, 24 MSA, 24 PSP | Features of normalized intensity uptake values of the ROIs (putamen, thalamus, anterior cingulate gyrus, pars opercularis) from 18F-DMFP-PET images | Feature selection: 2-sample t-test for the importance of each ROI; Classification: SVM + Bayesian network; Validation: tenfold cross-validation | Classification accuracy: PD vs. non-PD: SVM + Bayesian network (4 ROIs): 78.16% | Classification accuracy: PD vs. non-PD: Using all voxels: 70.11%; Using ROIs only in the striatum: 73.56%; SVM (major voting): 74.71%; Multiple-kernel SVM: 75.86% |
| Segovia et al., 2017a68 | 39 PD, 24 MSA, 24 PSP | Features of normalized intensity values of the ROIs in the caudate, putamen, thalamus, olfactory, and SMA from 18F-DMFP-PET images | Feature selection: 2-sample t-test for the importance of each ROI; Classification: Multiple-kernel-learning SVM; Validation: tenfold cross-validation | Classification accuracy: PD vs. non-PD: 73.56% (using 5 ROIs) | Using 5 ROIs, classification accuracy was higher than that using 2 ROIs in the striatum (68.96%); and higher than that using DATSCAN (59.77%) |
| Segovia et al., 2017b69 | 39 PD, 24 MSA, 24 PSP | Features of normalized intensity values of the ROIs in the striatum, which was automatically segmented from 18F-DMFP-PET images | Feature selection: none; Classification: SVM; Validation: fivefold cross-validation | Classification accuracy: PD vs. non-PD: Stratum using automatic segmentation: 75.86% | Classification accuracy: Stratum using atlas: 72.41%; All voxels: 65.52% |
| 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 |
| Glaab et al., 201921 | 44~60 PD, 14~16 HC | Whole-brain uptake data extracted from FDOPA PET and FDG-PET; Metabolomics data from blood plasma | Classification: SVM, random forest; Validation: Leave-one-out | SVM AUC for FDOPA + blood metabolomics: 0.98; SVM AUC for FDG + blood metabolomics: 0.91 | |
| Shen et al., 201922 | 125 PD, 225 HC | Uptake data of stratum and other regions extracted from FDG PET | Classification: GLS-DBN Validation: Train-validation ratio: 80:20 | Test set 1: Classification accuracy=90% (AUC = 0.912); Test set 2: Classification accuracy=86% (AUC = 0.899) | |
| Wu et al., 201923 | Cohort 1: 91 PD, 91 HC Cohort 2: 22 PD, 26 HC | Texture features of uptake data extracted from over 90 regions of interest on FDG PET using texture analysis | Classification: SVM Validation: fivefold cross-validation | Classification accuracy: Cohort 1: Accuracy = 91.26%; Cohort 2: Accuracy = 90.18% | |
| Zhao et al., 201924 | 502 PD, 239 MSA, 179 PSP | Saliency features (using saliency maps of regions of interests) of uptake data extracted from FDG PET | Classification: CNN Validation: sixfold cross-validation | Classification accuracy: For PD: Sensitivity = 97.7%, Specificity = 94.1%; For MSA: Sensitivity = 96.8%, Specificity = 99.5%; For PSP: Sensitivity = 83.3%, Specificity = 98.3% |
ADNI Alzheimer’s disease neuroimaging initiative, AUC area under the ROC (receiver-operating characteristic) curve, CADx computer-aided diagnosis, CIT or CT classification tree, CL caudate left, CNN convolutional neural networks, CR caudate right, CSF cerebrospinal fluid, EPNN enhanced probabilistic neural network, ET essential tremor, GLS-DBN group Lasso sparse deep belief network, GTL graph-based transductive learning, KNN k-nearest neighbor, LDA linear discriminant analysis, LR logistic regression, NM nearest mean, PCA principal component analysis, PD Parkinson’s disease, PDD pre-synaptic dopaminergic deficit, PL putamen left, PR putamen right, PLS partial least squares, PNN probabilistic neural network, PPMI Parkinson’s progression markers initiative, RBD rapid eye movement (REM) sleep behavior disorder, ROC receiver-operating characteristic, ROI region of interest, SVM support vector machine, SWEDD scans without evidence of dopaminergic deficit, VP vascular parkinsonism.