Participants
All data used in this study were obtained from the Parkinson’s Progression Markers Initiative (PPMI) database (www.ppmi-info.org/data) accessed on 3 April 2021. The dataset contained 790 pre-processed 123I-FP-CIT DAT-SPECT images acquired at the screening stage. In this study, we included cases acquired with the same manufacturer’s SPECT system (SIEMENSE) and were continuously enrolled in PPMI from the 790 cases. Table 1 shows the characteristics of the patients in this study. There were 291 patients in both the normal control (NC) and PD groups. The NC and PD group comprised 70 (49 men and 21 women, 61.2 ± 9.6 years) and 221 (140 men and 81 women, 60.0 ± 11.0 years) individuals, respectively. All data were divided into the training and test sets at 7:3 so that the ratio of the NC and PD groups would be constant. The number of the training and test sets for the NC and PD groups were 49 and 155 and 21 and 66, respectively.
Table 1 Patient characteristics
Characteristics
|
NC
|
PD
|
Number of patients
|
70
|
221
|
Male/female
|
49/21
|
140/81
|
Age
|
61.2 ± 9.6
|
60.0 ± 11.0
|
Age of diagnosis
|
NA
|
60.7 ± 9.5
|
Hoehn and Yahr stage
|
0.01 ± 0.1
|
1.58 ± 0.5
|
MDS-UPDRS Ⅲ
|
1.0 ± 1.9
|
21.8 ± 9.3
|
MoCA
|
28.2 ± 1.1
|
27.3 ± 2.3
|
MDS-UPDRS Movement Disorder Society-Unified Parkinson’s Disease Rating Scale, MoCA Montreal Cognitive Assessment, NC normal control, PD Parkinson’s disease, NA not applicable
Reconstruction and spatial normalisation of SPECT images
Reconstructed DAT-SPECT images were downloaded from the PPMI website. As per PPMI documentation, pre-processing steps were performed at the Institute for Neurodegenerative Disorders and included the following steps: SPECT imaging and reconstruction: SPECT imaging was acquired at each imaging centre as per the PPMI imaging protocol and sent to the institute for neurodegenerative disorders for processing. SPECT raw projection data were imported to a HERMES (Hermes Medical Solutions, Stockholm, Sweden) system for iterative reconstruction. Iterative reconstruction was performed without filtering. The reconstructed files were transferred to the PMOD (PMOD Technologies, Zurich, Switzerland) for subsequent processing. Attenuation correction ellipses were drawn on the images, and a Chang 0 attenuation correction was applied to images utilising a site-specific μ that was empirically derived from phantom data acquired during site initiation for the trial. Once attenuation correction was completed, a standard Gaussian three-dimensional (3D) 6.0-mm filter was applied.
Then, the DAT-SPECT images were spatially normalised to Montreal Neurologic Institute (MNI) space using statistical parametric mapping (SPM12, Wellcome Trust Centre for Neuroimaging, London, UK) in MATLAB R2021a (version 9.10, The MathWorks, Inc. Massachusetts, USA). DAT-SPECT images were spatially normalised to the MNI-based template of 123I-FP-CIT [27, 28] using the old normalise function under identical conditions. After spatial normalisation, the radiological technologist with 13-year clinical experience visually assessed for misalignment between DAT-SPECT and the template. The pre-processed images were saved in the Neuroimaging Informatics Technology Initiative format using 91 × 109 × 91 isotropic voxels of 2 mm.
Calculation of radiomics features and semi-quantitative indicators
The automated anatomical labelling atlas (AAL) 3 [29] voxel of interest (VOI) template was used to calculate the radiomics features. The feature calculation VOIs were the caudate nucleus, putamen, and pallidum (Fig. 1). The pallidum was added to the VOIs to compensate for errors in spatial normalisation. Radiomics features were calculated using Standardized Environment for Radiomics Analysis (SERA) [30–32] worked on MATLAB. One hundred eighty-six image biomarker standardisation initiative-standardised features [24] were calculated using SERA, including 50 first-order features (statistical, histogram, and intensity histogram features) and higher-order136 3D features (Table 2). A total of 558 radiomics features were calculated for the caudate, putamen, and pallidum VOIs. We also calculated the ratio of the caudate to the putamen or pallidum of radiomics features. These totalled 930 radiomics features. Furthermore, the SUR of the caudate nucleus (SURcaudate), putamen (SURputamen), and pallidum (SURpallidum) were calculated as conventional semi-quantification indices. The SUR was calculated using the following formula [33]:
![](data:image/png;base64,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)
where Cstriatum is the average count of the caudate nucleus, putamen, or pallidum and Cbackground is the average count of the occipital lobe. In addition, the ratios of the caudate to the putamen or pallidum (CRputamen, CRpallidum) were calculated. All the semi-quantitative indices were compared between the NC and PD groups, and receiver operating characteristic (ROC) [34] analysis was performed.
Table 2 Number of radiomics features per region and their family names
Feature family
|
Number of features
|
Local intensity
|
2
|
Intensity-based statistics
|
18
|
Intensity histogram
|
23
|
Intensity-volume histogram
|
7
|
GLCM
|
50
|
GLRLM
|
32
|
GLZSM
|
16
|
GLDZM
|
16
|
NGTDM
|
5
|
NGLDM
|
17
|
Total
|
186
|
GLCM grey-level co-occurrence matrix, GLRLM grey-level run-length matrix, GLZSM grey-level zone size matrix, GLDZM grey-level distance zone matrix, NGTDM neighbourhood grey tone difference matrix, NGLDM neighbourhood grey-level dependence matrix
Radiomics feature selection and signature construction
The least absolute shrinkage and selection operator (LASSO) [35] function in MATLAB was used to select effective features from the radiomics features. All radiomics features were z-scored to mean 0 and standard deviation 1.0 before being inputted to LASSO. LASSO permits the estimation and selection of explanatory variables [36, 37], that is, radiomics features with non-zero coefficients. For the selection of radiomics features using LASSO, a 10-fold cross-validation test was conducted using the training set. Furthermore, the linear combination sum of features with non-zero coefficients when the largest lambda value such that the mean square error (MSE) is within one standard error of the minimum MSE was used as the radiomics signature. We compared the classification performance of the radiomics signature and semi-quantitative indicator that showed the highest classification performance.
Classification model construction with radiomics signature and semi-quantitative indicator
The classification models for the NC and PD groups were constructed using the radiomics signature and/or semi-quantitative indicator. The four classifiers used were support vector machine (SVM), k-nearest neighbour (KNN), linear discriminant analysis (LDA), and decision tree. The main parameters of each classifier were as follows: SVM (BoxConstraint = 1, KernelScale = 1, KernelFunction = polynomial [order = 3]), KNN (NumNeighbors = 1, Distance = minkowski, Exponent = 2), LDA (Gamma = 0), and decision tree (MinLeafSize = 1, MinParentSize = 10). The features used were radiomics signature alone, semi-quantitative indicator alone, and both. The training set was used to train the classifier, and the performance of each classification model was evaluated using the test set. Classification performance was evaluated using the area under the ROC curve (AUC).
Statistical analyses
The radiomics signature and SURs in the NC and PD groups were tested for significant differences using the Wilcoxon rank-sum test. ROC analysis was performed using semi-quantitative indicators and radiomics signature. We used the DeLong [38] test to examine the differences in AUCs, and for multiple comparisons, the Bonferroni correction was performed. The sensitivity, specificity, and accuracy of semi-quantitative indices and radiomics signature were calculated using the optimal cut-off values determined based on ROC analysis. Differences were considered statistically significant at P < 0.05. All statistical analyses were performed using RStudio (version 1.4.1106).