Parkinson’s disease (PD) is a complex, progressive neurodegenerative disorder characterized by an insidious onset, high morbidity, and significant disability rates, posing a severe challenge to global public health systems. PD not only threatens the physical and mental health of middle-aged and elderly people but also markedly reduces their quality of life (1). The primary clinical manifestations of PD include motor symptoms (e.g., tremor, rigidity, bradykinesia, and postural instability) and nonmotor symptoms (e.g., sleep disturbances, olfactory dysfunction, autonomic dysfunction, cognitive impairment, and psychiatric disorders). According to the World Health Organization (2), the global number of PD patients exceeded 11.9 million in 2021 and is projected to rise to 25.2 million by 2050, representing a 112% increase. With increasing population ageing, the prevalence of PD is growing at an annual rate of 1.7% (3), imposing a substantial burden on patients, families, and society (4). PD can be classified into two major subtypes on the basis of clinical symptoms: tremor-dominant (TD) and postural instability–gait disturbance (PIGD) (5). The former is characterized by resting tremor (4–6 Hz), rigidity, and bradykinesia, often accompanied by cogwheel rigidity, whereas the latter manifests as balance impairment (positive Romberg sign), freezing of gait, and festination, which are frequently associated with autonomic dysfunction. These subtypes exhibit significant differences in dopaminergic drug responsiveness, disease progression rates, and nonmotor symptom profiles (6, 7). Therefore, developing a rapid and objective diagnostic method to distinguish these subtypes is a critical unmet need for improving patient outcomes.

Currently, PD subtyping relies heavily on subjective clinical assessments, which are prone to cognitive biases and variability in patient cooperation. As a result, clinicians often accurately diagnose PD subtypes only at advanced disease stages (5), delaying targeted treatment. Magnetic resonance imaging (MRI), a noninvasive, radiation-free, high-resolution imaging technique, is widely used for diagnosing and staging neurological disorders. Previous studies have identified structural alterations in the thalamus, caudate nucleus, putamen, and globus pallidus as key factors underlying TD and PIGD subtypes (8, 9). For example, Rosenberg-Katz et al. (10) reported that TD is more closely associated with thalamic damage, whereas PIGD is associated with more severe nigrostriatal degeneration. Dehghan et al. (11) reported widespread grey matter atrophy in PIGD patients, particularly in the frontal lobe, supplementary motor area (SMA), and basal ganglia. Bunzeck et al. (12) demonstrated differences in iron deposition between PD subtypes, with TD patients showing higher basal ganglia iron levels than PIGD patients. However, these studies focused primarily on macroscopic changes and failed to capture subtle structural variations, which limits their diagnostic accuracy and contributes to clinical misclassification. Such misdiagnosis may delay optimal treatment or exacerbate symptoms due to the use of inappropriate medication (13). Moreover, the lack of understanding of microstructural changes hinders the discovery of novel biomarkers.

Radiomics, introduced by Lambin et al. (14), is a computer-aided imaging technique that quantifies subtle, visually imperceptible features in medical images to assist in disease diagnosis and differentiation. It has been widely applied to PD, Alzheimer’s disease (AD), and brain tumours (1, 15, 16). However, existing radiomic models suffer from low accuracy and time-consuming workflows, limiting their clinical adoption (17). Prior to this study, we developed a convolutional neural network (CNN)-based AI model for automated segmentation and measurement of 109 brain regions, completing the process in under 30 s. This study leverages this network to rapidly segment brain nuclei, extract radiomic features, and construct machine learning models to differentiate TD and PIGD.

The study included 135 PD patients (43 PIGD, 92 TD). In the training set, 2,264 radiomics features were extracted from 8 deep brain nuclei. Automated technology precisely quantified the volumes of 109 brain subregions for each patient. For bilateral thalamic regions, the maximum relevance minimum redundancy (MRMR) algorithm initially selected 1,000 representative features from 4,528 radiomic features. Recursive feature elimination cross-validation (RFECV) further refined these to 17 optimal features. Similar feature selection pipelines were applied to the caudate nucleus, putamen, and globus pallidus, yielding region-specific optimal feature sets. Finally, features from all regions were integrated: correlation coefficients first selected 2,195 highly associated features from the full set, MRMR reduced these to 1,000 key features, and RFECV ultimately identified 17 highly discriminative combined features for model construction.

On the basis of single-region (thalamus, caudate, putamen, globus pallidus) and combined features, we constructed 45 prediction models using five algorithms (AdaBoost, BDT, GP, LR, RF), with detailed performance metrics shown in Table 1. Delong’s test revealed that the BDT model with combined features (5 pallidal, 4 putaminal, and 4 caudate features) performed best, achieving training and testing AUCs of 1.000 and 0.962, respectively (Figure 2), demonstrating excellent classification and generalization capabilities. Calibration curves revealed high agreement between the predicted and actual probabilities (Figure 3), whereas DCA demonstrated significant net clinical benefit across decision thresholds (Figure 4), providing a quantitative tool for personalized treatment decisions (see Table 2).

TD and PIGD differ not only in motor symptoms (resting tremor vs. gait freezing) but also in nonmotor symptom patterns (e.g., sleep disturbances, autonomic dysfunction) (19), leading to varied responses to PD medications (e.g., dopamine agonists vs. levodopa) (20, 21) and novel therapies such as transcranial magnetic stimulation. However, current PD subtyping relies on clinical differentiation, which is often delayed until intermediate or advanced stages (18, 22), resulting in late intervention. Thus, establishing early, rapid, and accurate subtyping methods could optimize symptom management and enable timely targeted interventions.

This study implemented automated whole-brain subregion segmentation using VB-NET, a U-NET variant. As an efficient medical image segmentation model, VB-NET automatically learns spatial features and boundary information of brain subregions without manual annotation, reducing the processing time from hours to seconds while improving precision and consistency. Previous segmentation methods have relied on manual delineation or simple algorithms. For example, Jingnan et al. (23) manually outlined the caudate and putamen to calculate the striatal-to-occipital uptake ratios (SORs) of 18F-DOPA in PD patients versus controls. These manual approaches are time-consuming, subjective, and inconsistent. Crucially, our model uses routine 3D-T1WI without advanced functional imaging, significantly reducing costs and enhancing clinical accessibility. Furthermore, VB-NET automatically extracted 2,264 radiomics features from key nuclei (the caudate, putamen, globus pallidus, and thalamus), capturing multidimensional morphological and textural information. Five machine learning algorithms (AB, BDT, GP, LR, and RF) were employed for automated TD/PIGD classification and were validated on independent test sets. The BDT classifier achieved exceptional performance (training AUC = 1.000; testing AUC = 0.962), demonstrating high accuracy and generalizability. Compared with prior studies, this work represents the first fully automated TD/PIGD classification pipeline, eliminating subjective clinical assessments and completing diagnosis within 3 min—offering clinicians a rapid, objective tool for personalized treatment planning.

Feature selection identified 17 core radiomic features with distinct regional distributions: caudate (8), globus pallidus (6), putamen (2), and thalamus (1). This pattern aligns with PD pathophysiology: pallidal (6) and putaminal (2) texture features showed early pathological (24) sensitivity, with abnormalities detectable in preclinical stages (25) (e.g., REM sleep behaviour disorder), preceding macroscopic atrophy. TD was strongly correlated with putaminal texture features (26) (e.g., autocorrelation). As a key node in the direct pathway, putamen dysfunction directly contributes to classic PD motor symptoms (tremor, movement initiation difficulty) (27). There exist distinct patterns of iron deposition in the basal ganglia between Parkinson’s disease (PD) patients with postural instability/gait disorder (PIGD) and those with tremor dominance (TD). Compared with the TD group, the PIGD group exhibits significantly higher susceptibility values in the left putamen (PUT), particularly in patients whose major symptomatic side is the right limb. This indicates that the distribution of iron deposition in the basal ganglia is more extensive in the PIGD group, involving not only the substantia nigra (SN) but also regions such as the caudate nucleus (CN) and putamen, with a more severe degree of iron deposition in the putamen (28). From the perspective of the correlation between iron deposition and symptoms, the damaged putamen is likely to be the pathophysiological basis for gait and posture disorders in Parkinson’s disease. Reduced autocorrelation indicates disrupted textural periodicity, which is consistent with PET evidence of decreased dopamine receptor density (29), reflecting early synaptic degeneration. PIGD involved primarily pallidal (5) and thalamic (1) features. The globus pallidus, a core indirect pathway node, exhibits degeneration-related motor inhibition (rigidity, bradykinesia) (30). Changes in texture homogeneity may reflect iron deposition (T2*/SWI hypointensity) or neuronal loss (31), whereas the presence of thalamic abnormalities suggest upstream motor circuit disruption (32). GrayLevelNonUniformity frequently appeared in the bilateral caudate (Caudate_R, Caudate_L, Caudate_R_boxmean), indicating microstructural heterogeneity from neuronal loss, iron deposition, or uneven Lewy body distribution. SmallAreaEmphasis in the right thalamus (Thalamus_R) and left globus pallidus (Pallidum_L) suggested early microstructural changes (e.g., enlarged perivascular spaces, focal gliosis), matching histopathological “nonuniform degeneration” patterns. These features not only enabled high-accuracy classification but also revealed spatiotemporal patterns of basal ganglia involvement in PD subtypes—early putamen/pallidum texture changes driving motor subtype differentiation, with thalamic features contributing to circuit-level dysfunction. Thus, these radiomic features provide quantifiable imaging biomarkers for precise subtyping and pathological investigation.

Our study had several limitations. This study had a retrospective cross-sectional design. Future prospective longitudinal studies should examine whether feature changes predict disease conversion (e.g., PD to dementia). The PD sample size was relatively small. All the data were acquired on 3 T Siemens scanners, but model generalizability requires validation across multicentre datasets. Future studies need to further verify the stability and universality of the conclusions of this study by expanding the sample size (including multi center and multi-regional data) and conducting independent external cohort validation. The current research results cannot directly verify their predictive efficacy for subtype progression trajectories (such as the rate of motor symptom deterioration, patterns of cognitive decline, etc.). In the future, we will conduct a longitudinal follow-up study based on a multi-center cohort to verify the longitudinal stability and clinical practical value of potential predictive biomarkers. The sole focus of this study was imaging features. Future work should integrate clinical scores (UPDRS, MMSE) into diagnostic tools. Although radiomic features demonstrated high accuracy in PD subtyping, their underlying pathological basis requires further elucidation.

We developed an AI model that accurately differentiates PD subtypes within 3 min using routine MRI data. This fully automated approach significantly improves diagnostic efficiency and holds substantial clinical potential for precise PD subtyping.