A review of FDG-PET biomarker analysis methods in Parkinson's disease

Abstract

Despite well-established clinical diagnostic criteria for Parkinson's disease (PD), challenges remain in early-stage diagnosis, differentiating atypical parkinsonism disorders, and understanding of underlying disease mechanisms. Neuroimaging, particularly positron emission tomography (PET) with [¹⁸F]fluorodeoxyglucose (FDG), plays a crucial role by offering in vivo assessments of brain activity through glucose metabolism, a marker of neuronal function. FDG-PET has demonstrated consistent metabolic abnormalities in PD, even in early or ambiguous stages, supporting its potential as both a diagnostic and mechanistic tool. Recent advancements in analytical methods, including regional univariate, pattern-based multivariate, network connectivity, and machine learning approaches, have allowed for more detailed exploration of PD-related brain dysfunction across multiple spatial, temporal, and organizational scales. This review provides an overview of the various FDG-PET analysis methods, discussing their principles, strengths, and limitations, with an emphasis on their utility for applications in PD, including both diagnostic applications and insights into underlying neural mechanisms.

Plain language summary title

A Review of FDG-PET Brain Imaging Analysis Methods in Parkinson's Disease

Plain language summary

Despite advances in diagnosing Parkinson's disease (PD), early detection and differentiation from similar disorders remain difficult. Brain imaging, such as FDG-PET, which provides a measure of how the brain uses glucose, can provide insight into brain activity and help address this challenge.

This review outlines the variety of approaches used to analyze FDG-PET brain scans to better understand PD. Some methods focus on individual brain regions to identify where activity is higher or lower than normal. Others look at patterns across the whole-brain, capturing how multiple regions change together. More complex approaches study how different parts of the brain are connected, revealing disruptions in large-scale brain networks. Machine learning methods using computer algorithms can identify patterns that can help classify individuals or predict disease progression.

By comparing these approaches, this review highlights how FDG-PET imaging can support diagnosis and improve understanding of how PD affects the brain. It also helps guide researchers in selecting the most appropriate analytical method depending on the clinical question.

Introduction

Parkinson's disease (PD) is a progressive neurodegenerative disorder characterized by a broad range of motor and non-motor symptoms, reflecting widespread dysfunction across cortical and subcortical brain systems. Despite well-established clinical diagnostic criteria, accurate diagnosis remains challenging, particularly in early disease stages and when distinguishing PD from atypical parkinsonism disorders (APDs). ¹ Moreover, clinical measures alone provide limited insight into underlying disease mechanisms, progression, and treatment responses. ²

While advances in α-synuclein seeding amplification assays (SAAs) have enhanced detection of pathology and informed biomarker-based PD classification, these techniques still do not capture the neurodegenerative processes and large-scale brain dysfunction driving clinical symptoms and progression and do not reliably distinguish between APD subtypes.^3,4 Neuroimaging therefore plays a critical complementary role, providing in vivo measures of structural and functional disease-related brain changes that can support differential diagnosis, reveal mechanisms of disease progression, and serve as quantitative biomarkers for evaluating potential disease-modifying interventions. ²

Among the available imaging modalities, positron emission tomography (PET) with [¹⁸F]fluorodeoxyglucose (FDG) has emerged as particularly well suited for PD biomarker development. FDG-PET is deemed to provide a quantitative measure of glucose metabolism, tightly related to neuronal activity, offering a window into the functional state of brain circuits. In PD, FDG-PET has consistently demonstrated reproducible metabolic abnormalities, including changes that may be detectable even in early or clinically ambiguous stages, positioning it as a valuable tool for both diagnosis and elucidating disease-related brain mechanisms.^5–7

Beyond identifying regional metabolic abnormalities, FDG-PET data lend themselves to a wide range of analytical approaches that probe brain function at multiple spatial, temporal, and organizational scales. This is particularly important in PD, where pathological changes are not confined to a single brain region but instead involve distributed neural networks. Although early studies have focused on regional univariate analyses that quantify metabolic differences within predefined regions of interest (ROIs), subsequent methodological advances have introduced multivariate pattern-based techniques capable of capturing distributed disease-related metabolic signatures. More recently, network connectivity analyses and machine learning approaches have leveraged FDG-PET data to model inter-regional relationships, characterize network topology, and perform subject-level group classification. Together, these multiscale analytical capabilities are particularly relevant for PD biomarker development because of the disease's heterogeneous symptoms and widespread brain involvement.

The goal of this review is to provide a narrative methodological overview of FDG-PET-based biomarker analysis methods, with a focus on the diverse range of methodological approaches that have been implemented thus far, rather than on detailed study results and clinical interpretations. Representative studies were selected based on their relevance to the methodological concepts discussed, illustrating the major analytical frameworks and their applications, rather than to provide a systematic or exhaustive survey of the literature.

We organize existing approaches into four broad categories: regional univariate analyses, pattern-based multivariate analyses, network connectivity analyses, and machine learning approaches. For each category, we discuss methodological principles and key strengths and limitations, with an emphasis on their utility for applications in PD, including both diagnostic applications and insights into underlying neural mechanisms. Figure 1 summarizes the methodological categories used throughout this review.

Figure 1.

Conceptual overview of the four major categories of FDG-PET analysis methods in PD reviewed in this article. Abbreviations: HC, healthy control; SSM-PCA, scaled subprofile modeling-principal component analysis; PLS, partial least squares; CCA, canonical correlation analysis; IRCA, inter-regional correlation analysis; ICA, independent component analysis; PC, principal component; LR, logistic regression; SVM, support vector machine; RF, random forest; CNN, convolutional neural network; TPR, true positive rate; FPR, false positive rate.

Regional univariate analyses

Which brain regions exhibit abnormal metabolism in PD?

Regional univariate analyses examine metabolism within predefined ROIs or across individual voxels, treating each ROI or voxel independently. Group differences or associations with clinical measures are typically assessed using t-tests or correlation analyses. The main advantage of this approach is interpretability: changes in discrete, anatomically defined regions can be easily interpreted in relation to PD pathology, making it particularly useful when prior knowledge points to a specific target.

Early FDG-PET studies in PD identified hypometabolism in the occipital and parietal cortices, while metabolism within the striatum was relatively preserved despite marked nigrostriatal dopaminergic degeneration.^8,9 More recent hypothesis-driven studies have reported striatal hypermetabolism in PD relative to atypical parkinsonian disorders (APDs), ¹⁰ as well as frontal hypometabolism along electrode trajectories following subthalamic nucleus deep brain stimulation. ¹¹ Voxelwise FDG-PET univariate analyses have also demonstrated utility in improving diagnostic accuracy when distinguishing PD from APDs such as multiple system atrophy (MSA), progressive supranuclear palsy (PSP), corticobasal syndrome (CBS), and dementia with Lewy bodies (DLB).^12,13

While this strategy provides important insights into disease-related changes, advances in neuroscience have increasingly highlighted that the brain operates as an interconnected network rather than as a collection of isolated regions. ¹⁴ Because altered metabolism in one region is likely to have widespread effects across the brain, univariate analyses may therefore be ill-posed to fully capture the nature of metabolic dysfunction in PD. Beyond this conceptual limitation, statistical constraints further limit these methods, including susceptibility to multiple comparisons and reduced sensitivity to distributed or heterogeneous disease effects. Despite these limitations, univariate analyses provide an anatomically intuitive window into disease-related metabolic changes, supporting mechanistic interpretation and contributing to diagnostic efforts.

Pattern-based multivariate analyses

What set of metabolic alterations characterize PD?

Unlike univariate approaches, pattern-based multivariate analyses model metabolism across multiple brain regions simultaneously to identify disease-specific spatial covariance patterns. This approach offers three key advantages: (i) captures coordinated metabolic changes across distributed brain regions, (ii) can increase sensitivity by leveraging shared variance across the brain, and (iii) reduces the multiple comparison burden inherent to univariate testing. This makes this approach especially well-suited for capturing subtle, distributed metabolic network changes characteristic of PD. However, a key trade-off is interpretability: while univariate changes are anatomically localized, multivariate patterns are inherently distributed and often less directly tied to a single structure.

Scaled subprofile modeling-principal component analysis (SSM-PCA)

The scaled subprofile model using principal component analysis (SSM-PCA) is a well-established multivariate technique for deriving spatial patterns that summarize disease-related variation.^15,16 In this framework, SSM removes subject- and group-level global effects to focus on disease-related signals in the data, after which unsupervised PCA (without the use of diagnostic group labels) decomposes the data into two outputs: (i) orthogonal principal components (PCs), which are spatial patterns capturing the dominant sources of variance across subjects, and (ii) subject scores, which quantify the expression of each pattern in the individual subjects.

Typically, a set of ROIs or voxels spanning the entire brain is used, allowing the extracted components to reflect distributed, whole-brain metabolic patterns. Interpretation of these patterns is guided by the subject scores describing the strength with which a subject expresses a specific pattern. Patterns whose scores discriminate between healthy and diseased groups and/or correlate with external clinical variables are inferred to represent disease-related metabolic signatures.

In the context of PD, SSM-PCA has been used extensively to identify the Parkinson's disease-related pattern (PDRP) and Parkinson's disease cognitive pattern (PDCP), which reflect characteristic alterations in metabolic activity across cortical and subcortical regions.^17–21 These patterns have been shown to be highly reproducible,^22–29 and have proven useful not only for quantifying disease severity, progression, and response to intervention,^{18,23,30–36} but also for identifying APD-specific metabolic patterns that enable differentiation from PD,^25,37–41 and for demonstrating that individuals with idiopathic REM sleep behavior disorder (iRBD) exhibit metabolic patterns resembling the PDRP, suggesting the presence of a prodromal metabolic phase preceding clinical PD diagnosis. ⁴² Together, these findings support SSM-PCA as a robust FDG-PET biomarker for characterizing disease severity, progression, and treatment-related changes in PD. Comprehensive reviews of FDG-PET SSM-PCA as a clinical biomarker in PD are provided by Gu et al. ⁴³ and Dhawan et al. ⁴⁴

Supervised extensions of SSM-PCA

Logistic regression (LR), canonical variates analysis (CVA), linear regression, and ordinal trends CVA (OrT-CVA) can all be used as supervised extensions of SSM-PCA that combine multiple orthogonal PCs derived from the original analysis.^45,46 LR and CVA both optimize the linear combination of components to maximize group separation, with CVA explicitly maximizing between-group variance relative to within-group variance and LR optimizing a likelihood-based objective. Alternatively, linear regression can be utilized to quantify how strongly each component relates to a clinical measure, enabling the selection and combination of only those PCs that correlate with the measure of interest. OrT-CVA differs by explicitly constraining the derived pattern to change monotonically within each subject across a set of ordered conditions (e.g., longitudinal disease timepoints or systematically varied experimental parameters), thereby identifying spatial patterns that exhibit consistent within-subject directional trends.

These approaches are particularly useful for deriving FDG-PET metabolic patterns optimized to capture different dimensions of PD. For example, Meles et al. applied LR to SSM-PCA components to generate a weighted PDRP that maximally discriminated between PD patients and healthy controls (HC). ²⁹ Matthews et al. applied CVA to construct a supervised PDRP, then derived a separate age-related metabolic pattern and assessed the spatial overlap between the two, revealing partial overlap between age-related and disease-related effects. ⁴⁷ Linear regression has similarly been applied to derive the PDCP, yielding patterns that correlate with cognitive performance.^20,26 Mure et al. used OrT-CVA to identify a tremor-related metabolic pattern (PDTP) by modulating deep brain stimulation parameters within individual subjects and enforcing ordinal constraints on tremor severity. ⁴⁸

Partial least squares (PLS) and canonical correlation analysis (CCA)

Partial least squares (PLS) and canonical correlation analysis (CCA) provide multivariate frameworks for identifying spatial patterns of brain metabolism in relation to secondary datasets.^49–51 In both approaches, paired latent variables are derived that capture shared structure between the imaging data and external variables (maximizing covariance in PLS and correlation in CCA).

In PD applications, these external variables are typically clinical measures of disease, leading to a key advantage of PLS and CCA: the resulting spatial patterns are intrinsically linked to clinical measures rather than emerging solely from variance within the imaging data. Additionally, multiple clinical variables can be incorporated within a single model, enabling identification of the most relevant disease dimensions in a data-driven exploratory manner. Bevington et al. applied PLS to identify FDG-PET metabolic covariance patterns associated with distinct aspects of PD and aging, revealing insights beyond those of unsupervised or single-clinical-dimension approaches. ⁵² This method has also been applied to identify metabolic patterns underlying visual hallucinations in PD ⁵³ and to link genetic data with combined FDG-PET and resting-state functional magnetic resonance imaging (rs-fMRI) measures in PD. ⁵⁴

Dynamic mode decomposition (DMD)

Dynamic mode decomposition (DMD) is a method that identifies spatial patterns associated with temporal dynamics, decomposing ordered data into spatiotemporal modes. ⁵⁵ This enables the capture of spatial brain patterns which have complex, potentially oscillatory or non-monotonic disease trajectories.

Although DMD has not yet been applied to FDG-PET data to our knowledge, its potential relevance to PD has been demonstrated in related PET modalities. For example, Fu et al. applied DMD to multi-tracer PET data to identify spatiotemporal patterns of putaminal dopamine processing across disease stages, revealing dynamic alterations that were not readily captured by static or monotonic progression analyses.^56,57 This work illustrates how DMD can uncover latent disease-related dynamics by jointly modeling spatial expression and temporal evolution, allowing the identification of spatial patterns with independent temporal behaviors that may correspond to different underlying disease mechanisms.

Network connectivity analyses

How are different brain regions metabolically connected in PD, and how does network organization differ due to disease?

PD is increasingly understood as a network disorder, in which pathological processes manifest not only as focal regional abnormalities but also as disruptions of inter-regional interactions across the brain. ⁵⁸ In the context of FDG-PET, metabolic connectivity is typically defined as the statistical association between glucose metabolism in different brain regions, reflecting coordinated metabolic activity at the systems level, a concept that has been explored since the 1980s.^59–61 Network connectivity analyses leverage this concept to characterize large-scale metabolic organization by quantifying how metabolism in one region covaries with metabolism elsewhere.

In contrast to pattern-based approaches, which capture spatial covariance across regions without modeling direct inter-regional interactions, network methods explicitly quantify these interdependencies and the brain's topological organization, offering insight into the distributed mechanisms underlying PD and its clinical features. Importantly, converging evidence from functional neuroimaging suggests that such network-level abnormalities may arise early in the disease course, including during preclinical and prodromal stages, making connectivity analyses particularly relevant for understanding early pathophysiological changes and for biomarker development.^62,63 However, these approaches tend to yield more abstract representations that are sometimes less directly interpretable biologically and typically require larger sample sizes for reliable estimation.^64,65

Inter-regional correlation analysis (IRCA)

Inter-regional correlation analysis (IRCA) defines metabolic connectivity as the Pearson correlation between glucose uptake measured in two regions, calculated either across subjects in static PET (sPET) datasets, ⁶⁶ or across time within a subject in functional PET (fPET) acquisitions. ⁶⁷ In both cases, high correlation is interpreted as evidence of coordinated metabolic activity between regions. This approach is conceptually straightforward, allowing researchers to examine how metabolism in a seed region relates to the rest of the brain.

An important conceptual distinction arises between sPET- and fPET-based implementations of metabolic connectivity. In fPET, correlations are computed across time within an individual subject's data and therefore reflect moment-to-moment coupling of metabolic fluctuations, more closely paralleling the interpretation of functional connectivity in rs-fMRI. However, the reliability of these time-resolved correlations depends on signal quality, and fPET estimates may be particularly sensitive to noise when computed over short temporal frames. ⁶⁸ Additionally, as fPET remains a relatively recent and rapidly evolving methodology, acquisition and analysis protocols have yet to become standardized. Existing studies differ substantially in tracer administration strategies, including bolus injection, constant infusion, and bolus-plus-infusion approaches, as well as in temporal binning, complicating direct comparisons across studies and centers. Continued efforts toward harmonizing acquisition and analysis protocols will be important for improving reproducibility and facilitating multicenter validation of fPET-derived biomarkers.

In contrast, although sPET acquisition protocols are comparatively well established, correlations are computed across subjects, such that “connectivity” reflects consistent patterns of inter-regional metabolic covariation across individuals rather than dynamic coupling within the brain. The extent to which these two formulations identify the same underlying network organization depends on assumptions of ergodicity: that inter-individual variability mirrors intra-individual temporal dynamics.^69,70 As a result, the biological interpretation of metabolic connectivity derived from sPET remains less well defined, highlighting the importance of careful distinction between cross-sectional and within-subject formulations when drawing inferences about metabolic network organization in PD.

Thus far, IRCA has been predominantly applied to PD and related disorders using cross-sectional sPET datasets, where inter-regional correlations across subjects have revealed systematic disruptions in large-scale brain systems. Sala et al. demonstrated reduced connectivity within and between default mode, sensorimotor, and associative networks in DLB, consistent with widespread network vulnerability beyond regional metabolism changes. ⁷¹ Verger et al. extended this framework by linking metabolic connectivity patterns to dopaminergic pathways. ⁷² Caminiti et al. further characterized alterations in metabolic connectivity architecture, revealing sex-specific effects and distinct network signatures associated with early- versus late-onset disease, and showing that these metabolic connectivity alterations spatially correlate with several neurotransmitter systems beyond dopamine.^73,74

Recently, fPET-based implementations have begun to extend IRCA toward within-subject analyses of metabolic dynamics. In PD, M. Li et al. applied network-level correlation analyses to FDG-PET data to characterize changes in network topology across disease stages. ⁷⁵ Heinecke et al. used fPET to map time-resolved metabolic coupling, demonstrating altered dynamic interactions within motor and associative networks in PD. ⁷⁶ However, a remaining key limitation of IRCA is that it considers each pairwise correlation in isolation, making estimates susceptible to indirect effects and motivating the use of multivariate network models that jointly estimate connectivity structure.

Partial correlation connectivity

Partial correlation approaches extend IRCA by estimating the direct relationship between two brain regions while controlling for the influence of all other regions in the network. ⁷⁷ In high-dimensional FDG-PET datasets, where the number of regions often exceeds the number of subjects, sparse inverse covariance estimation (SICE) or graphical lasso (GLASSO) is commonly used to stabilize these estimates.^77,78 By controlling for indirect effects, partial correlation connectivity can more accurately isolate direct interactions, providing a clearer view of network structure. However, the introduction of sparsity constraints requires careful selection of hyperparameters (such as the sparsity level that determines the number of retained connections), which influence the resulting connectivity estimates and affect interpretability. ⁷⁹

In PD and related disorders, partial correlation methods using FDG-PET have revealed widespread alterations in metabolic connectivity^80–82 supporting the idea that PD is a “networkopathy” with widespread connectivity alterations rather than isolated focal changes. ⁵⁸ In particular, studies have reported decreased connectivity in frontal and cerebellar regions and increased connectivity in basal ganglia structures, indicating compensatory or pathological network reconfiguration.^83,84 Additionally, they have highlighted that connectivity alterations align with dopaminergic-vulnerable pathways, suggesting a mechanistic link between disease processes and network disruption.^80–82 See Sala et al., ⁶⁴ Hanania et al., ⁶⁵ and Horowitz et al. ⁸⁵ for detailed reviews on metabolic connectivity methodology and application to PD.

Independent component analysis (ICA)

Independent component analysis (ICA) can be used to identify statistically independent metabolic networks without requiring predefined seed regions. ⁶¹ It can be applied either across subjects (sPET) or across time within a subject (fPET), capturing large-scale networks such as the default mode, sensorimotor, and cognitive control systems. Unlike IRCA or partial correlation approaches, ICA defines networks as multivariate entities rather than as collections of pairwise relationships. However, ICA components are generally more abstract and less directly interpretable, potentially requiring cross-validation with other modalities (e.g., rs-fMRI) to provide functional context. ⁶¹ Reliable estimation also typically demands larger sample sizes compared to simpler connectivity measures. ⁸⁶

In PD, ICA has revealed alterations in several resting-state metabolic networks such as the default mode network, illustrating that the method can serve as a powerful tool for capturing distributed, multivariate network alterations in PD, complementing seed-based and partial correlation approaches.^87–89

Joint sparse structural estimation (JSSE)

Joint sparse structural estimation (JSSE) quantifies metabolic connectivity by measuring the similarity between voxelwise distributions of glucose uptake between brain regions. ⁹⁰ For each subject, voxel intensities within an ROI are treated as a probability distribution, and connectivity between two regions is defined using the Jensen-Shannon (JS) divergence between their respective distributions. Regions with more similar metabolic distributions are thus inferred to be more strongly connected. Importantly, this notion of connectivity reflects distributional similarity rather than direct physiological coupling, making JSSE a relatively indirect and abstract measure of metabolic network organization. As a result, JSSE-derived networks can be difficult to interpret biologically and may be particularly sensitive to ROI selection as well as reconstruction and preprocessing choices.^90–92

Despite these limitations, a key advantage of JSSE is its ability to generate fully subject-specific metabolic networks from standard sPET data. In PD, JSSE-based approaches have been used to characterize alterations in individual-level metabolic network topology, including disrupted connectivity between network hubs and abnormal motor-limbic connectivity patterns.^90,91 Additionally, JSSE connectivity has been shown as a promising diagnostic feature when paired with machine learning classification, outperforming regional FDG-PET and rs-fMRI connectivity measures in distinguishing PD from HCs, suggesting its potential as a complementary biomarker.^92,93

Graph theory metrics

Graph theory provides a formal framework to quantify the topological organization of metabolic connectivity networks. ⁹⁴ Nodes are defined as brain regions, and edges represent connections derived from IRCA, partial correlation, ICA, or JSSE. A wide range of graph-theory metrics such as degree, clustering coefficient, path length, and centrality can be calculated, summarizing both local and global network properties, offering insights into integration, segregation, efficiency, and hub structure.

In PD, graph-theory analyses of FDG-PET-derived metabolic networks have consistently revealed alterations in large-scale network organization, further supporting the view of PD as a network disorder rather than a collection of isolated regional abnormalities.^80,91,93,95 These studies frequently report reorganization of hub structure and connectivity involving subcortical regions, alongside altered network efficiency and clustering.

While graph-theory metrics offer a powerful abstraction for summarizing complex metabolic networks, their interpretation can be challenging, as many metrics lack a direct biological correspondence and multiple measures are often tested simultaneously. ⁹⁶ Additionally, different studies often focus on different metrics and use different network construction methods, which can complicate comparisons across the literature. For a comprehensive overview of graph-theory approaches applied to FDG-PET metabolic connectivity, see Tuan et al. ⁹⁷

Machine learning approaches

Can computational models classify individuals or predict clinical outcomes in PD based on complex metabolic features?

Machine learning (ML) refers to computational methods that learn patterns and relationships from data without explicit programming. ML techniques are particularly suited for FDG-PET analysis in PD, where metabolic brain patterns are high-dimensional, spatially distributed, and subtle. ⁹⁸ While traditional methods may miss distributed or non-linear patterns, ML can exploit the full multivariate structure of FDG-PET data, enhancing diagnostic discrimination and prognostic modeling. However, this often comes at the cost of interpretability. ⁹⁹

Most ML pipelines follow two stages: feature extraction/selection and supervised classification. Raw voxelwise FDG-PET data are often transformed into a reduced feature space by either limiting voxels to a hypothesis-based ROI, or using methods such as SSM-PCA. Alternatively, radiomics approaches have also been used, which involve extraction of a large number of quantitative features capturing characteristics such as intensity, shape, texture, and spatial relationships. When the number of features exceeds the number of subjects, additional dimensionality reduction is often applied to retain the most informative variables and mitigate overfitting. The final stage of the pipeline is classification, in which ML algorithms learn to map the extracted features onto an outcome of interest, such as distinguishing PD from HCs or identifying APDs.

Traditional machine learning (ML) techniques

Traditional ML algorithms such as the decision tree (DT), random forest (RF), logistic regression (LR), support vector machine (SVM), and Generalized matrix learning vector quantization (GMLVQ) are common methods of group classification. These algorithms take the selected features as input and learn to predict either a binary outcome (PD vs. HCs) or one of multiple classes (distinguishing APD subtypes) using labeled training data. Table 1 gives a brief summary of these methods. See Binson et al. for a complete overview of ML algorithms for biomedical applications. ¹⁰⁰

Table 1.

Traditional ML techniques.

Method	Description
Decision Trees (DT)	Non-linear tree structures that recursively split features to separate classes; intuitive but prone to overfitting.101,102
Random Forests (RF)	Ensembles of DTs trained on random subsets of features; reduces overfitting through averaging. 102
Logistic Regression (LR)	Linear model predicting probability of outcomes using a weighted sum of features; interpretable but limited in capturing non-linear patterns.103,104
Support Vector Machines (SVM)	Finds hyperplanes that maximize the margin between classes; typically uses the “kernel trick” for non-linear separations; less interpretable.105,106
Generalized Matrix Learning Vector Quantization (GMLVQ)	Learns representative prototypes for each class and a relevance matrix, providing interpretable class representations and feature relevance. 107

These ML approaches have been widely applied to classify PD and to differentiate between distinct APDs using FDG-PET data. They differ primarily by the types of input features, the choice of classifier, and the included patient sub-groups. Collectively, these studies have shown reliable detection of disease-specific metabolic features associated with PD and APDs, while also showing potential for prognostic applications, such as the prediction of progression from iRBD to PD and from PD to PD with dementia (PDD).

Table 2 provides a summary of representative studies employing these ML classifiers for PD and APD differentiation using FDG-PET data. See S. Peng et al. for further discussion of the challenges, limitations, and future directions of ML classification for FDG-PET PD data. ¹⁰⁸

Table 2.

Summary of studies using ML classifiers for PD classification with FDG-PET data.

Study	Input features	Classifier method	Patient groups included
Tang et al. 109	SSM-PCA	LR	PD, MSA, PSP
Garraux et al. 110	Voxelwise: whole-brain	SVM	PD, MSA, PSP, CBS
Mudali et al. 111	SSM-PCA	DT	PD, MSA, PSP, HC
Tripathi et al. 112	SSM-PCA	LR	PD, MSA, PSP
van Veen et al. 113	SSM-PCA	SVM and GMLVQ	PD, AD, HC
Y. Wu et al. 114	Radiomic features	SVM and RF	PD, HC
Glaab et al. 115	Voxelwise: whole-brain, along with blood metabolomics	SVM and RF	PD, HC
Rus et al. 116	SSM-PCA	LR	PD, MSA, PSP, HC
van Veen et al. 117	SSM-PCA	GMLVQ	PD, HC
Hu et al. 118	Radiomic features from subcortical areas, along with other MRI-based structural features	LR	PD, MSA
Schindlbeck et al. 119	SSM-PCA	LR	PD, MSA, PSP
Booth et al. 120	Voxelwise: whole-brain	SVM	PD-mild cognitive impairment, PDD
van Veen et al. 121	SSM-PCA	GMLVQ	HC, PD, DLB, AD, iRBD
W. Li et al. 90	JSSE	SVM	PD, HC
Chen et al. 92	JSSE	RF	PD, HC
Ruppert-Junck et al. 122	Voxelwise: midbrain	RF	PD, HC
X. Li et al. 93	JSSE with rs-fMRI	SVM	PD, HC

Deep learning (DL) techniques

While the previously discussed classifier methods have been shown to be effective in many contexts, they rely on manual feature extraction and selection. This approach works well when there is a clear hypothesis or predefined structure to the data. However, when such a priori structure is difficult or impossible to predict, and as the classification problem becomes more complex and the dimensionality of the data increases, traditional methods may struggle to capture and select all the relevant patterns. In such cases, highly complex feature engineering would be required, limiting the practicality of fully utilizing the data's potential.

This is where deep learning (DL) techniques, particularly convolutional neural networks (CNNs), can offer a powerful alternative. ¹²³ CNNs are specifically designed to work directly with raw, high-dimensional data such as voxelwise FDG-PET images. ¹²⁴ In contrast to fully connected layers, which connect every input to every output, convolutional layers apply local filters across the input, allowing the network to automatically detect spatial or temporal patterns while dramatically reducing the number of parameters. Rather than relying on manual feature extraction, CNNs learn hierarchical patterns directly from the raw input during training, allowing them to automatically select for the intricate, complex features needed to make accurate classifications.

Once the CNN has extracted and selected the relevant features, the output can be fed into a traditional classifier, such as those discussed above. However, more commonly, the classifier is directly incorporated into the CNN architecture itself through a final fully connected layer enabling an end-to-end learning process.

While CNNs excel at learning localized spatial features from FDG-PET data, they are inherently limited in capturing long-range dependencies across the brain's metabolic network due to the localized nature of convolutional filters. This limitation has motivated the exploration of transformer-based deep learning models, which leverage self-attention mechanisms to capture relationships between distant elements in the input. By complementing CNN local feature extraction with transformer global dependency modeling, hybrid approaches have emerged as a promising direction for improving PD classification performance.

Despite their advantages in automating feature learning and classification, DL methods also present several challenges. They typically require large amounts of labeled data for training, substantial computational resources, and careful tuning of a large number of hyperparameters (e.g., learning rate, network depth, filter sizes, batch size, and regularization settings). ¹²⁵ Additionally, given the complexity of their internal structure, DL models are often considered a “black box”, making them less interpretable than traditional machine learning methods. ¹²⁶

DL-based approaches using FDG-PET data have been increasingly applied to classify PD and differentiate APDs, with several studies reporting high diagnostic performance. ⁹⁸ Table 3 provides a summary of representative studies employing DL models for PD and APD differentiation using FDG-PET data. See Wang et al., ¹²⁷ T. Zhao et al., ¹²⁸ and Sadabad et al. ⁹⁸ for detailed systematic reviews and meta-analyses on DL-based approaches for PD classification.

Table 3.

Summary of studies using DL models for PD classification with FDG-PET data.

Study	Input data	DL model	Patient groups included
Manzanera et al. 129	Voxelwise: PDRP relevant regions	CNN with fully connected layer	PD vs HC
Y. Zhao et al. 130	Voxelwise: whole-brain	CNN with fully connected layer	PD, MSA, PSP
Choi et al. 131	Voxelwise: whole-brain	CNN with fully connected layer	AD, PD, HC
Dai et al. 132	Voxelwise: substantia nigra, fusion image of FDG-PET and structural MRI	CNN with fully connected layer	PD vs HC
Sun et al. 133	Voxelwise: whole-brain	CNN with SVM	PD, HC
P. Wu et al. 134	Voxelwise: whole-brain	CNN with fully connected layer	PD, MSA, PSP, HC
Zhu et al. 135	Voxelwise: whole-brain	CNN with SVM	PD non-freezing-of-gait, PD freezing-of-gait
S. Wu et al. 136	Voxelwise: whole-brain	CNN and transformer with fully connected layer	PD, MSA, PSP
C. Li et al. 137	Voxelwise: whole-brain with radiomic features	CNN and transformer with XGBoost classifier (ensemble DT method)	PD, MSA, PSP

Summary of methods

The following table (Table 4) summarizes the discussed methods used to analyze metabolic FDG-PET data, highlighting their primary application, advantages, and potential limitations to guide interpretation and methodological choices.

Table 4.

Summary of metabolic imaging analysis methods.

Method	Application	Advantages	Limitations
Regional Univariate Analyses	Compare metabolism in individual regions	High anatomical interpretability; simple to implement	Limited sensitivity to distributed or network-level effects; ignores inter-regional relationships
SSM-PCA	Identify dominant metabolic covariance patterns	Captures distributed patterns across the whole brain; subject-specific pattern expression scores can be readily computed for independent datasets; widely validated	Less anatomically intuitive; relevance is established post hoc
PLS and CCA	Identify metabolic patterns maximally associated with clinical or external variables	Enables exploratory identification of imaging-clinical relationships and cross-modal relationships	Not widely validated; latent variable interpretation can be complex
IRCA	Estimate inter-regional metabolic connectivity	Captures metabolic relationships between regions; simple to implement	Captures both direct and indirect connections; sensitive to noise in fPET; biological interpretation in sPET can be complex
Partial Correlation Connectivity	Estimate direct metabolic relationships controlling for indirect effects	Isolates direct connections for a clearer network structure	Sensitive to model assumptions; fPET and sPET limitations
ICA	Identify statistically independent metabolic connectivity networks	Captures large-scale networks without requiring predefined seed regions	May lack biological interpretability without validation; larger datasets are typically required; fPET and sPET limitations
Graph Theory Metrics	Quantify topological properties of metabolic networks	Provides a compact framework to summarize complex properties of networks	Metrics can be abstract with difficult biological interpretation; results depend on network construction method; fPET and sPET limitations
Traditional ML Techniques	Perform supervised classification using extracted features	Enables subject-level classification from a set of features; often allows for some model interpretability	Requires feature engineering; risk of overfitting in small datasets
DL Techniques	Perform supervised classification directly from raw FDG-PET data	Provides automatic feature learning from high-dimensional raw images	Requires large datasets to limit risk of overfitting; computationally intensive; very limited interpretability

Discussion

FDG-PET provides a powerful approach for investigating functional alterations in PD, offering a quantitative measure of brain glucose metabolism that reflects large-scale changes in neural systems. This is particularly important in PD, where pathological changes are not confined to a single brain region but instead involve distributed neural networks. As reviewed here, analytical methods applied to FDG-PET data span a broad methodological spectrum, from regional univariate analyses to multivariate pattern modeling, network connectivity approaches, and modern ML frameworks. The uniqueness of each approach has been leveraged to capture different aspects of disease-related metabolic organization, where together they provide complementary perspectives that enhance our ability to diagnose and characterize PD, as well as monitor therapeutic intervention.

Despite advances in FDG-PET biomarker methodologies, several important considerations remain. Fundamentally, FDG-PET does not directly measure the underlying molecular pathology of PD, such as α-synuclein aggregation or degeneration of dopaminergic and other neurotransmitter systems. Rather, alterations in FDG-PET uptake likely reflect the functional consequences of these pathological processes, providing an indirect measure of disease-related brain dysfunction. It is the distributed nature of these metabolic abnormalities across multiple neural systems that has motivated the development of multivariate pattern analysis, connectivity methods, and ML approaches that characterize distributed metabolic relationships across the brain rather than isolated regional changes.

Additionally, although FDG-PET is often interpreted as a proxy for neuronal activity, FDG uptake reflects glucose utilization across multiple cell types, including glial cells, which contribute substantially to FDG uptake even under normal physiological conditions.^138,139 This distinction is particularly relevant in PD, where baseline signal can be further modulated by neuroinflammation mediated by activated glial cells.^139,140 Moreover, FDG measures realized glucose metabolism rather than metabolic demand, which may be disconnected from underlying neuronal activity in disease-states where energy supplies are impaired. Separately, longitudinal FDG-PET assessments of disease progression may be confounded by symptomatic therapies, including pharmacologic and surgical interventions, which can alter regional metabolism independently of underlying neurodegeneration.^33,35 Taken together, these considerations indicate that FDG-PET findings should be interpreted with caution and highlight the advantage of multimodal imaging approaches that integrate metabolic, structural, and functional measures to better characterize various aspects of brain function.

Beyond these biological considerations, methodological variability in image acquisition, preprocessing, and analytical pipelines remains an important challenge for quantitative neuroimaging biomarker development. Intensity normalization strategies, spatial normalization, image smoothing, and ROI or atlas definitions can influence quantitative metabolic estimates, contributing to heterogeneity in reported findings and complicating comparisons across studies. ¹⁴¹ Multicenter studies have shown that these technical differences can substantially affect pattern expression scores and the performance of supervised classification methods.^113,117,142 Consequently, although center-related variation is expected to affect all the analytical approaches discussed in this review, it is likely to be particularly problematic for finer-grained applications, such as trials of disease-modifying therapy or studies of disease progression, where subject or group-specific biological or treatment-related differences are subtler than those between PD and HCs. To improve reproducibility and facilitate translation into clinical practice, future multicenter studies should prioritize standardized acquisition, reconstruction, and preprocessing protocols whenever possible. When complete standardization is not feasible, image harmonization methods represent an active area of research for reducing scanner- and reconstruction-related differences while preserving quantitative accuracy.^143,144

Many advanced analytical approaches also require relatively large datasets and careful validation to ensure robustness and avoid overfitting, particularly for DL models. While these models may achieve strong predictive performance, they often sacrifice interpretability, making it difficult to relate learned representations directly to underlying neurobiological mechanisms. Additional practical constraints, including data privacy regulations and restricted data sharing between institutions, can further limit the development and external validation of large-scale predictive models in clinical populations.^127,145 Notably, whereas some methods like SSM-PCA have been validated across multiple populations, many emerging approaches still lack independent validation cohorts, emphasizing the need for external replication to confirm generalizability. Emerging techniques such as federated learning aim to address these challenges by enabling model training across multiple sites without requiring the sharing of raw data. ¹⁴⁵

Conclusion

Overall, the diversity of analytical frameworks applied to FDG-PET highlights the richness of metabolic information available for studying PD. Continued methodological development, along with larger and more standardized datasets and well-validated analysis pipelines, will be essential for translating these approaches into robust biomarkers. Such advances show promise to improve diagnosis, enable tracking of disease progression, and deepen understanding of the network-level mechanisms underlying PD.