Early Detection of Parkinson’s Disease Using Automatic Classification of Single-photon Emission Computed Tomography Images Authors: Jihad Boucherouite, Abdelilah Jilbab, Atman Jbari Published: 2026 ABSTRACT AbstractObjectivesParkinson’s disease (PD) is a chronic neurodegenerative disorder characterized by central nervous system dysfunction. Early identification may enable prompt treatment and help slow the progression of disabling symptoms. Previous studies have reported that clinical assessments based on visual interpretation may be insufficiently accurate and often miss early PD. Therefore, this study aimed to develop an automated single-photon emission computed tomography-based model for binary classification of healthy controls and patients with early PD.MethodsWe analyzed 514 DaTSCAN images from the Parkinson’s Progression Markers Initiative database, using one unique scan per individual. The workflow comprised three main stages: image processing; computation of 23 features, including radial, threshold, boundary, and striatal binding ratio features; and image classification using machine learning algorithms.ResultsThe medium Gaussian support vector machine achieved an accuracy of 97.09% ± 1.53% (95% confidence interval, 95.11%–99.07%), sensitivity of 98.18% ± 1.48%, specificity of 93.85% ± 3.44%, and area under the receiver operating characteristic curve of 98.49% ± 1.31%. This performance was significantly higher than that of the three-dimensional convolutional neural network baseline model (accuracy, 94.55% ± 2.98%; p = 0.045), while requiring substantially less training time.ConclusionsIn this dataset, carefully designed feature engineering combined with classical machine learning outperformed the deep-learning baseline when the training data were limited and the selected features aligned with clinical diagnostic criteria. This approach achieved high accuracy for early PD detection and may provide computational efficiency and interpretability suitable for clinical implementation.Keywords: Parkinson Disease, Tomography, Early Diagnosis, Image Processing, Computer-Assisted, Support Vector Machine Introduction Parkinson’s disease (PD) is a chronic neurodegenerative disorder that affects the central nervous system. Its symptoms include tremor, bradykinesia, freezing of gait, postural instability, postural deformity, and rigidity [1]. PD primarily involves degeneration of dopaminergic neurons in the striatum, including the caudate and putamen, resulting in dopamine depletion [2]. Brain imaging modalities used in PD assessment include magnetic resonance imaging, functional magnetic resonance imaging, positron emission tomography, and single-photon emission computed tomography (SPECT) [3]. Positron emission tomography and SPECT are particularly useful for detecting PD at an early stage [4]. SPECT is commonly used in European clinics for early PD diagnosis [2] and can help distinguish PD from drug-induced parkinsonism, multiple system atrophy, essential tremor, and progressive supranuclear palsy [5]. However, early-stage PD remains difficult to diagnose and requires specialized expertise. Moreover, the accuracy of clinical diagnosis has remained limited over the past 25 years, with pooled accuracy not exceeding 82.7% [6]. Various machine learning (ML) approaches have been explored for early PD detection [7], and automated image analysis methods may support more objective diagnosis. Image processing of brain scans is a promising approach because it enables biomarkers to be extracted directly from SPECT images. Staff et al. [8] and Palumbo et al. [9] classified DaTSCAN images as PD or healthy control (HC) by segmenting the striatum, applying shape analysis, and using probabilistic neural networks, classification trees, Naive Bayes methods, or singular value decomposition of voxels in the striatal region; these approaches achieved accuracy of up to 94.8%. Castillo-Barnes et al. [10] used striatal voxel intensity values as support vector machine (SVM) classifier features and obtained an accuracy of 97.04%. Segovia et al. [11] analyzed the right and left striatum separately using partial least squares to evaluate the maximum information encoded in the striatal region, achieving an accuracy of 94.7%. Several studies have also used feature engineering approaches, in which manually or semi-automatically extracted features provide interpretable markers of striatal degeneration. Prashanth et al. [12] used striatal binding ratios (SBRs) for both caudate nuclei and putamina and incorporated shapebased features as inputs to a diagnostic model, whereas Bhalchandra et al. [13] proposed radial and gradient features. In recent years, deep-learning models have also been applied to DaTSCAN SPECT imaging [14,15]. These models enable automated feature extraction from raw images but typically require larger datasets and offer limited interpretability. Feature engineering, by contrast, preserves explicit clinical meaning while requiring fewer samples; therefore, the present work sought to retain interpretability while automating the feature extraction pipeline. Building on our previous work [16], this study aimed to develop an automated SPECT image analysis model to support early objective diagnosis of PD, reduce observer variability, and aid clinical decision-making. Using SPECT images from the Parkinson’s Progression Markers Initiative (PPMI) database, the workflow comprised image processing; extraction of radial, threshold, boundary, and SBR features; and ML classification. Methods This study aimed to distinguish patients with early PD from HCs. Figure 1 shows the analysis workflow. First, SPECT images from the PPMI database were processed for slice selection. Second, boundary, radial, and threshold features were calculated. Finally, ML classification was performed. Flow diagram of the study analysis. PPMI: Parkinson’s Progression Markers Initiative, SPECT: single-photon emission computed tomography, SBR: striatal binding ratio. Dataset Image Preprocessing The PPMI database [17] includes patients with PD from different countries and at different stages of disease. The dataset used in this study comprised 514 unique individuals categorized into two groups: 385 patients with PD (251 male and 134 female) and 129 HCs (89 male and 40 female). Because the database contains longitudinal data with multiple scans per participant, one unique scan per individual was used to avoid data leakage [18]. The demographic distribution is presented in Table 1. The three-dimensional (3D) SPECT scans first underwent preprocessing of the raw projection data. Chang 0 attenuation correction was applied, followed by a standard 3D Gaussian filter with a radius of 6.0 mm. Final normalization ensured that all scans were anatomically aligned with the Montreal Neurological Institute (MNI) standard coordinate system [19]. The preprocessed SPECT images consisted of 91 two-dimensional slices, each measuring 109 × 91 pixels. Slices 39–42 were selected because the putamina and caudate nuclei were clearly visible in this range; the slice with the largest aggregated striatal surface area was then retained for further analysis. This selection was based on methods used in previous studies [16,20], careful examination of the striatal region, and Society of Nuclear Medicine (SNM) recommendations, which specify (i) use of at least three consecutive slices in the region of interest and (ii) consistency in slice selection. Table 1Demographic and clinical characteristics of the study cohortCharacteristicHealthy controlsParkinson’s diseaseNumber of subjects129385 Male89251 Female40134 Age (yr)61.27 ± 11.1061.45 ± 9.65 Hoehn & Yahr stage Stage I (unilateral involvement)-181 (47.0) Stage II (bilateral involvement)-201 (52.2) Stage not reported-3 (0.8)Values are presented as mean ± standard deviation or number (%).. Demographic and clinical characteristics of the study cohort Values are presented as mean ± standard deviation or number (%).. Image Processing Bhalchandra et al. [13] transformed SPECT images into binary images using a fixed threshold in the range of 0.8–0.9, selected by observing the highest activity in the caudate nuclei and putamina. The effectiveness of these preprocessing steps was evaluated by radiologists and neurologists. Prashanth et al. [12] used binarization with a fixed threshold selected through experimentation and observation of high-uptake regions. Rumman et al. [20] segmented SPECT images using unsharp masking, adaptive thresholding (0.52–0.71), and edge detection with a sequential grassfire algorithm. In the present study, several image-processing procedures were applied to improve visual quality and facilitate subsequent analysis. The first procedure was designed to reduce noise, enhance structural detail, and emphasize the striatal region. Each slice was first converted from RGB to grayscale, resized by a factor of two to increase spatial clarity, and sharpened using a Gaussian kernel with a width of 4 and a strength of 2. The unsharp-masking operation was computed according to Equation (1), following the formulation described in [21]. where y(n, m) is the sharpened image, x(n, m) is the input image, and λ denotes the strength of the sharpening effect and serves as a positive scaling factor that controls the degree of contrast enhancement in the output image. In this study, λ was set to 2 because preliminary experiments indicated that this value provided the most stable and visually consistent enhancement. z(n, m) is the correction signal computed as the output of a linear high-pass filter. The second procedure used thresholding to segment the striatum from the extracted slices and select the appropriate slice for each patient. The main challenge in grayscale images is determining a threshold that enables slice segmentation and subsequent calculation of radial and boundary features. Previous studies used several approaches: one study relied on an expert physician to localize the striatal region [19], another used the average of bimodal image intensity values from the histogram as the threshold [22], and two studies selected a value from a predetermined range based on a fixed standard deviation [12,13]. In this study, a fully automated process was developed using a global thresholding algorithm that converges on the interclass difference between the main region and the background [23]. This approach is based on images with a bimodally distributed histogram and produces a personalized threshold, T. Figure 2 shows the histogram of a PD SPECT image. The initial threshold, T, was selected by calculating the weighted mean intensity of all pixels, after which the image was segmented using T. This process divided the image pixels into two regions. F1 comprised pixels with grayscale values >T, and F2 consisted of pixels with grayscale values ≤T. Average intensity was then calculated for both regions. The threshold T was updated as the mean of both averages, as shown in Equation (2), and the process was repeated until the threshold value no longer changed. Histogram of a Parkinson’s disease single-photon emission computed tomography (SPECT) slice. The third procedure applied Canny edge detection [13], a de facto method for detecting object boundaries in images based on intensity gradients, to extract shape-descriptive features. Feature Extraction and Classification Models For each SPECT image, three types of features were extracted: boundary features (10-dimensional), radial features (8-dimensional), and SBR values (4-dimensional). The threshold feature T (1-dimensional), extracted during image processing, was then added, resulting in a 23-dimensional feature vector. 1) Boundary features Boundary features were computed separately for the left and right striatal regions. These features included area, circularity, major axis length, equivalent diameter, and perimeter, yielding 10 values in total (5 features × 2 sides). The features were calculated using the following equations: Area (A): number of pixels within the striatal region × (pixel spacing)2. Area (A): number of pixels within the striatal region × (pixel spacing)2. Perimeter (P): total boundary length in pixels × pixel spacing. Perimeter (P): total boundary length in pixels × pixel spacing. Circularity (C): C = 4πA/P2 where P is the perimeter. Circularity (C): C = 4πA/P2 where P is the perimeter. Major axis length: length of the longest line passing through the region’s centroid. Major axis length: length of the longest line passing through the region’s centroid. Equivalent diameter Deq:Deq=4A/π Equivalent diameter Deq:Deq=4A/π A uniform pixel spacing of 2.0 mm was applied to all MNI-normalized SPECT images. 2) Radial features In addition to boundary features, radial features were extracted to capture shape irregularities in the striatal regions [24,13]. Each radial measure represented the distance from the striatal centroid to the boundary along a predefined angular direction. Specifically, eight radi values were computed at 45° intervals, starting from the horizontal axis (0°) and rotating clockwise. where C(xc, yc) is the centroid of the striatal region and Bi(xi, yi) is the boundary point at angle θi = 45° × (i – 1). For spatial consistency, the right striatum was horizontally flipped before computation so that corresponding angular directions aligned across hemispheres. The final feature vector was obtained by averaging the right and left radi values for each angular direction: 3) SBR features SBR quantifies specific striatal radiotracer binding in the caudate nuclei and putamina of both hemispheres relative to the occipital cortex. The PPMI SPECT laboratory team extracted values from these striatal regions using a standardized template. SBR values were computed according to Equation (5), as detailed in [19]: The data file contained a four-dimensional vector representing the corresponding SBR measures. SBR values for all PPMI images were extracted from the PPMI database. In this study, each SBR measurement was assigned to the closest available SPECT image, minimizing the temporal difference between the two acquisitions. To ensure temporal consistency between imaging data and SBR measurements, the time difference was consistently quantified as Δt = |AgeSBR – AgeScan | in years. The median Δt was 0.10 years, or approximately 36 days; 95.9% of scans were paired within 0.25 years, or approximately 3 months, and 99.6% were paired within 1 year. These values support the temporal proximity of the assigned SBRs and imaging sessions, reducing potential bias due to longitudinal disease progression. Combining quantitative SBR measures with imaging data for early PD detection is consistent with established practices in the field [12,25]. Because the striatal dopaminergic signal declines by approximately 5%–12% annually in early PD [26], the expected SBR change within the median temporal window of 36 days, with 95.9% of pairs within 3 months, is <3%, which is within measurement precision and supports the validity of this strategy. 4) Model classification This study implemented 24 classifiers using the classification learner application, including SVM, naive Bayes, logistic regression, decision tree, discriminant analysis, k-nearest neighbor (KNN), and ensemble classifiers. Table 2 provides an overview of the hyperparameter settings for the best-performing classifiers. A 3D convolutional neural network (CNN) was also implemented, comprising three convolutional blocks with progressively increasing feature maps (8 → 16 → 32 filters). Each block contained a 3 × 3 × 3 convolutional layer, rectified linear unit activation, and 2 × 2 × 2 max-pooling with a stride of 2. The model was trained using the Adam optimizer with cross-entropy loss for 50 epochs per fold. All experiments used five-fold cross-validation and were executed on an Intel Core i7-10750H CPU at 2.60 GHz without graphics processing unit acceleration. Table 2Hyperparameter settings of the top-performing models for binary classificationModelHyperparametersMedium Gaussian SVMKernel: Gaussian;Kernel Scale: 4.4;Box Constraint (C): 1;Standardization: Yes;Multiclass coding: one-vs-one. Linear SVMKernel: Linear;Kernel Scale: Automatic;Box Constraint (C): 1;Standardization: Yes;Multiclass coding: one-vs-one. Cosine KNNNumber of neighbors: 10;Distance metric: Cosine;Distance weight: Equal;Standardization: Yes;Multiclass coding: one-vs-one. Cubic KNNNumber of neighbors: 10;Distance metric: minkowski (Cubi)c;Distance weight: Equal;Standardization: Yes;Multiclass coding: one-vs-one. RUSBoosted trees ensembleLearner type: decision tree;Maximum number of splits: 20;Number of learners: 30;Learning rate: 0.1;Number of predictors to sample: select all.SVM: support vector machine, KNN: k-nearest neighbor. Hyperparameter settings of the top-performing models for binary classification SVM: support vector machine, KNN: k-nearest neighbor. The performance metrics for the classifiers were computed as accuracy, sensitivity, and specificity. These metrics were calculated using the following equations: where TP (true positives) indicates the number of patients with PD who were correctly classified. TN (true negatives) indicates the number of HCs who were correctly classified. FP (false positives) indicates the number of HCs who were incorrectly classified as having PD. FN (false negatives) indicates the number of patients with PD who were incorrectly classified as HC. Results This section presents the results of image processing, feature computation, and classification. Figure 3 shows the segmentation results for three patients using a fixed threshold of 0.7 (Figure 3B) and global thresholding (Figure 3C). Table 3 shows the distribution and variation in area and circularity for four patients with PD and four HCs. With global thresholding, HCs showed greater circularity and area than with fixed thresholding, whereas patients with PD showed smaller values, enabling better discrimination between PD and HC. Fixed and global thresholding applied in three patients: (A) original grayscale image, (B) fixed threshold of 0.7, and (C) global threshold for each patient. Table 3Variation in area and circularity in four patients with PD and four HCsSubject numberThreshold fixeGlobal thresholdCategory AreaCircularityAreaCircularity306912557.2019582.04HC 307117780.0220391.39HC 307218084.4821493.12HC 307321484.6823391.57HC 301212450.8710342.95PD 301413257.2410644.49PD 301811845.498634.89PD 302016971.589944.99PDPD: Parkinson’s disease, HC: healthy control. Variation in area and circularity in four patients with PD and four HCs PD: Parkinson’s disease, HC: healthy control. To assess the discriminative power of the extracted features, statistical analyses were performed comparing PD and HC. Table 4 presents the mean and standard deviation of all calculated features. Threshold values derived by global thresholding differed significantly between groups (p < 0.01), with threshold values of 0.79 ± 0.04 for PD and 0.71 ± 0.12 for HC, indicating tight clustering around the mean in each group. All boundary features differed significantly between groups (p < 0.01) and were lower in PD than in HC. Most radial features, including rad2, rad3, rad4, rad6, rad7, and rad8, differed significantly between groups (p < 0.01) and were lower in PD, whereas rad1 and rad5 did not show this pattern. The complete feature extraction pipeline required 175.38 seconds. Table 4Boundary, radial, SBR, and global threshold features (presented as means with SDs) for healthy controls and patients with early PDFeaturesHCPDp-value MeanSDMeanSDArea, left105.6946.7868.6028.751.4910e-24 Circularity, left45.5916.3929.8512.265.9340e-28 Major axis length, left24.908.3519.654.851.9632e-17 Equivalent diameter, left11.382.019.171.803.6763e-27 Perimeter, left92.7840.8361.2227.137.6518e-22 Area, right81.6832.6655.3223.003.4859e-22 Circularity, right36.6214.4924.2110.045.9416e-25 Major axis length, right20.588.2417.526.692.3005e-05 Equivalent diameter, right9.623.397.962.651.4372e-08 Perimeter, right72.0228.7849.2520.382.5183e-21 rad119.0214.9417.469.010.1542 rad223.2213.4915.805.435.1508e-18 rad328.7212.0715.915.138.3095e-5 rad424.2113.4814.974.303.3200e-29 rad518.9015.5317.409.050.1781 rad622.519.3517.104.674.2480e-17 rad728.206.1816.404.624.9079e-82 rad822.708.3416.454.351.2831e-25 SBR left caudate2.950.622.000.564.0296e-48 SBR right caudate2.890.592.000.581.9494e-43 SBR left putamen2.080.560.820.355.5117e-119 SBR right putamen2.100.580.850.362.7885e-113 Global threshold0.710.120.790.046.1190e-30SBR: striatal binding ratio, HC: health control, PD: Parkinson’s disease, SD: standard deviation. Boundary, radial, SBR, and global threshold features (presented as means with SDs) for healthy controls and patients with early PD SBR: striatal binding ratio, HC: health control, PD: Parkinson’s disease, SD: standard deviation. Table 5 presents a performance comparison across 24 classical ML classifiers and the 3D CNN, with all metrics reported as mean ± standard deviation across folds. The medium Gaussian SVM achieved the highest performance, with an accuracy of 97.09% ± 1.53% (95% confidence interval: 95.11%–99.07%), sensitivity of 98.18% ± 1.48%, specificity of 93.85% ± 3.44%, and an area under the receiver operating characteristic curve (AUC) of 98.49% ± 1.31%. The low standard deviation indicated consistent performance across folds. The misclassification cost was 15, the prediction speed was approximately 27,000 observations/second, and the training time was 2.73 seconds. Figure 4 illustrates the confusion matrix (Figure 4A) and receiver operating characteristic curve (Figure 4B). Several other ML models also achieved competitive performance. The linear SVM obtained an accuracy of 96.30% ± 0.81%, whereas the cosine KNN and cubic KNN achieved accuracies of 96.69% ± 1.49% and 96.30% ± 0.83%, respectively. Ensemble methods, particularly the RUSBoosted trees ensemble (96.30% ± 0.81%) and bagged trees ensemble (95.91% ± 2.51%), also demonstrated strong performance. Table 5Five-fold cross-validation performance comparison across all classifiersModelAccuracy (%)Sensitivity (%)Specificity (%)AUCTraining Time (s)Model size (kB)Fold 1Fold 2Fold 3Fold 4Fold 5Mean ± SDLinear SVM96.0895.1597.0996.1297.0996.30 ± 0.8197.66 ± 1.6992.18 ± 6.2898.79 ± 0.760.89319Quadratic SVM96.0895.1597.0994.1795.1595.53 ± 1.1097.92 ± 1.9788.34 ± 2.9497.98 ± 1.180.97417Cubic SVM95.1094.1796.1294.1794.1794.75 ± 0.8696.62 ± 1.7489.14 ± 5.0497.15 ± 2.071.73921Fine Gaussian SVM86.2790.2987.3891.2692.2389.49 ± 2.5598.96 ± 1.0961.11 ± 11.9797.71 ± 1.680.83773Medium Gaussian SVM 99.02 97.09 95.15 96.12 98.06 97.09 ± 1.53 98.18 ± 1.48 93.85 ± 3.44 98.49 ± 1.31 2.527 25 Coarse Gaussian SVM99.0294.1798.0693.2092.2395.34 ± 3.0297.92 ± 2.6987.66 ± 8.7398.50 ± 0.851.20932Gaussian naive Bayes92.1688.3585.4485.4488.3587.95 ± 2.7792.99 ± 3.7472.98 ± 0.7295.79 ± 1.851.12813Kernel naive Bayes95.1094.1786.4192.2388.3591.25 ± 3.7593.51 ± 1.8484.55 ± 10.8296.41 ± 2.013.317436Fine tree93.1498.0691.2697.0990.2993.97 ± 3.4694.29 ± 4.5593.02 ± 5.7294.32 ± 3.400.93110Medium tree96.0893.2096.1294.1795.1594.94 ± 1.2695.32 ± 1.7493.78 ± 5.8594.99 ± 2.772.5610Coarse tree96.0892.2392.2397.0995.1594.56 ± 2.2396.10 ± 2.0589.91 ± 8.8894.40 ± 5.384.2687Linear discriminant98.0497.0994.1798.0689.3295.34 ± 3.7297.66 ± 1.6988.40 ± 11.1998.47 ± 1.730.87514Quadratic discriminant92.1696.1296.1297.0994.1795.13 ± 1.9796.88 ± 2.3689.88 ± 6.5797.04 ± 1.550.84124Fine KNN95.1096.1292.2396.1295.1594.94 ± 1.5996.10 ± 2.0591.42 ± 6.4993.76 ± 2.880.774102Medium KNN99.0295.1594.1798.0695.1596.31 ± 2.1097.14 ± 2.1393.82 ± 2.0798.25 ± 1.761.386102Coarse KNN88.2489.3294.1793.2097.0992.40 ± 3.6397.92 ± 0.7175.85 ± 13.3298.25 ± 0.853.056102Cosine KNN95.1095.1598.0697.0998.0696.69 ± 1.4997.40 ± 2.4394.58 ± 2.0898.37 ± 0.961.257102Cubic KNN95.1096.1297.0997.0996.1296.30 ± 0.8397.40 ± 1.5993.05 ± 1.6598.30 ± 1.250.796102Weighted KNN96.0895.1595.1597.0996.1295.91 ± 0.8197.14 ± 1.0992.28 ± 4.6898.43 ± 1.171.384102Boosted trees ensemble75.4974.7693.2096.1274.7682.87 ± 10.8298.70 ± 2.2535.38 ± 48.4564.63 ± 25.93316Bagged trees ensemble96.0899.0395.1592.2397.0995.91 ± 2.5197.66 ± 2.5090.68 ± 5.2298.01 ± 1.403.101296Subspace discriminant ensemble95.1096.1297.0994.1794.1795.33 ± 1.2797.66 ± 1.6988.37 ± 4.7298.66 ± 0.742.749220Subspace KNN ensemble91.1894.1792.2387.3894.1791.83 ± 2.8094.81 ± 0.7582.83 ± 12.7096.15 ± 1.933.1582000RUSBoosted trees ensemble96.0896.1295.1597.0997.0996.30 ± 0.8196.62 ± 0.7195.32 ± 3.2998.17 ± 0.583.677313 3D CNN 97.09 91.26 98.06 94.17 92.16 94.55 ± 2.98 96.90 ± 3.00 87.60 ± 8.30 98.80 ± 0.70 12376.4 12646 SVM: support vector machine, KNN: k-nearest neighbour, SD: standard deviation, AUC: area under the receiver operating characteristic curve.The bold font indicates the performance metrics of the highest-performing classifier. Five-fold cross-validation performance comparison across all classifiers SVM: support vector machine, KNN: k-nearest neighbour, SD: standard deviation, AUC: area under the receiver operating characteristic curve. The bold font indicates the performance metrics of the highest-performing classifier. Confusion matrix (A) and receiver operating characteristic curve (B) for the medium Gaussian support vector machine (SVM). To compare this approach with a representative deep learning method, a 3D CNN was implemented to operate directly on the full volumetric SPECT scans. The model achieved an accuracy of 94.55% ± 2.98%, sensitivity of 96.90% ± 3.00%, specificity of 87.60% ± 8.30%, and AUC of 98.80% ± 0.70%. Although the 3D CNN achieved competitive overall performance and a similarly high area under the curve, several performance differences were observed (Figure 5). Its accuracy varied more across folds than that of the medium Gaussian SVM, as reflected by the higher standard deviation (2.98% vs. 1.53%). Its specificity was also lower and less stable (87.60% ± 8.30% vs. 93.85% ± 3.44%). Training time was considerably longer, at 206 minutes. A paired t-test applied to fold-wise accuracies showed that the medium Gaussian SVM significantly outperformed the 3D CNN (p = 0.045), suggesting a consistent advantage for the classical model despite the use of a modern deep learning architecture. Performance comparison between the medium Gaussian support vector machine (SVM) and the three-dimensional convolutional neural network (3D CNN). Discussion This study presents an automated SPECT-based classification system for early PD detection that integrates image processing, feature extraction, and ML. The approach combined boundary, radial, and SBR features with adaptive global thresholding and achieved diagnostic accuracy of 97.09%, which was higher than the 94.55% accuracy of the 3D CNN baseline. Compared with the 3D CNN, the feature engineering approach produced higher and more stable accuracy, required fewer computational resources, remained robust with limited training data, and provided greater interpretability through clinically meaningful anatomical and functional features. This approach performed comparably to previously reported feature engineering and deep-learning methods. Prashanth et al. [12] manually selected the segmentation threshold for each patient and obtained an accuracy of 97.29% using SVM classifiers; the mean ± standard deviation values of the selected thresholds for HC and PD were 0.63 ± 0.04 and 0.69 ± 0.05, respectively. In a related study, Prashanth et al. [27] used only SBR features in SVM classifiers with a radial basis function kernel and obtained an accuracy of 96%. The present model achieved numerically higher performance than several existing methods that used other techniques, including CNNs [28], feed-forward neural networks [29], and artificial neural networks [20]. Magesh et al. [30] reported 95.2% accuracy using a VGG16-based model with transfer learning and incorporated local interpretable model-agnostic explanations to enhance interpretability. However, these studies used different datasets, database sizes, PD stages, and classification methods, which limits direct comparison. Future research will prioritize external validation using independent, multi-institutional datasets to assess generalizability across diverse clinical settings. The classification framework will also be extended from binary to multiclass classification to distinguish HCs, patients with Hoehn and Yahr stage 1 PD, and patients with Hoehn and Yahr stage 2 PD, thereby enabling disease severity assessment. In addition, future work will incorporate scans without evidence of dopaminergic deficit in patients with a clinical diagnosis of PD. Complementary feature types will also be explored to improve diagnostic accuracy. Footnotes Conflict of Interest No potential conflict of interest relevant to this article was reported.