A curriculum-guided unified framework for robust unsupervised domain adaptation on multi-cohort Parkinson’s disease diagnosis

Introduction

As one of the most common neurodegenerative disorders, Parkinson’s disease (PD) impacts over millions of individuals globally and, without available curative interventions, progressively results in severe functional impairment (1). Pharmacologic treatments can effectively palliate symptoms particularly in the early stage, making accurate and timely diagnosis crucial for clinical management (2). Current diagnostic approaches depend predominantly on observable motor symptoms, for example through walking video (3), and patient history (4), typically when clinical features have already progressed past the optimal intervention period. Haruyama et al. (3) utilize feature extraction and classifier technology to analyze walking videos, which provide cost-effective, non-invasive screening tools based on motor symptoms, but they capture phenotypic manifestations rather than the underlying neuropathological changes. This diagnostic challenge results in frequent early-stage misidentification, with accuracy rates based on established guidelines varying between 73.8% to 83.9%, significantly influenced by clinical expertise (5,6). Such constraints have motivated researchers to identify reliable and sensitive markers capable of recognizing PD during prodromal or initial phases. Medical imaging techniques, particularly magnetic resonance imaging (MRI), demonstrate significant potential in identifying minor cerebral changes that precede symptomatic presentation (7). For instance, diffusion MRI and structural MRI can collectively capture both macro- and microstructural alterations in PD, thereby enhancing early detection capabilities (8,9).

Despite MRI’s potential for early PD detection, conventional interpretation of neuroimaging data demands substantial expertise and suffers from subjective variability. The emergence of deep learning technologies has revolutionized this field by facilitating automated and unbiased feature extraction from imaging data (10). Neural network architectures have achieved remarkable diagnostic performance for PD identification using neuroimaging modalities (11). Both T1-weighted and diffusion MRI sequences have become standard inputs for deep learning-based neurodegenerative disease classification. Diagnostic models utilizing T1W MRI have reached accuracy level exceeding 70%, while diffusion MRI-based approaches demonstrate even higher performance with over 77% accuracy (12). Three-dimensional (3D) residual network (ResNet) (13), VGG16 (14), Inception V3 (15), AlexNet (16), simple fully convolutional neural (SFCN) network (12), Transformer model (17) and the variants of these network (18) have shown relevant findings in the diagnosis of PD.

Most PD diagnosis models are developed and validated on the Parkinson’s Progression Markers Initiative (PPMI) database, which is a multi-site cohort with unified scanning protocols and extensive data. These high-performance models trained on large-scale data with standardized protocols hold tremendous potential for clinical applications across medical institutions. While deep learning models trained on PPMI achieve excellent diagnostic performance, these models exhibit significant performance degradation when applied to external MRI datasets from different medical institutions. This decline in performance is often attributed to domain shift (19) caused by variations in MRI protocols, scanner manufacturers, acquisition parameters, and population characteristics between source and target domains.

Consequently, cross-domain inconsistency not only limits the applicability of individual models, but even more critically, impedes the entire translation process of PD AI diagnostic systems from research to clinical practice. Reliable and generalizable automated diagnosis is critical for early and accurate PD detection, yet current models’ domain dependency severely undermines their practical value in real clinical scenarios. Therefore, developing robust domain adaptation (DA) approaches that ensure the transferability of deep learning models across heterogeneous clinical cohorts is essential for realizing equitable and scalable AI-assisted PD diagnostic tools.

Unsupervised domain adaptation (UDA) (20) requires only unlabeled data from the target domain and labeled data from the source domain, emerging as a promising solution to multi-center PD diagnosis. The core objective of UDA is to learn domain-invariant features that minimize the discrepancy between the source and target distributions, thereby enabling a model trained on labeled source data to generalize effectively to the unlabeled target domain. Discrepancy-based, adversarial-based and self-training methods are representative UDA methodologies. Discrepancy-based methods explicitly measure and minimize a statistical distance between the source and target feature representations through designed different loss metrics. Adversarial-based methods employ a domain discriminator to adversarially train the feature extractor to produce features that are indistinguishable between source and target domains. Discrepancy loss and adversarial training are often implemented together to align the feature distributions of the two domains within the deep network. Self-training methods leverage model generated pseudo-labels on the target domain as supervisory signals for iterative training, and pseudo labels often are obtained by selecting high-confidence predictions. The combination of self-training and adversarial training can result in more distinguishable feature learning model and generate more accurate pseudo labels for self-training, which better comprehensively utilizes the advantage of both of them and increasingly becomes the mainstream method.

However, the integration of adversarial training with self-training in UDA faces significant challenges, primarily concerning the acquisition of noisy pseudo-labels in the target domain, which leads to error propagation and model instability. And the confidence threshold for pseudo-label selection mainly depends on manual setting and is sensitivity to hyperparameters, further complicates the effective harmonization of these two strategies. Furthermore, the conflicting objectives of domain-invariant feature learning and pseudo-label-based discriminative learning often result in suboptimal alignment and performance degradation. Targeting the above problems, curriculum learning (21) that generates dynamic threshold and adaptively selects more accurate pseudo label according to current model training performance is introduced into this study. In this way, we propose to introduce curriculum learning guided pseudo label generation into self-training adversarial UDA framework and constrain the prediction consistency of masked target data and original target data according to generated pseudo label for better feature representation learning. Experiments on multi-cohort PD MRI data demonstrate that our proposed algorithm achieves the outperformed classification performance on different source and target domain data, compared to state-of-the-art discrepancy-based, adversarial-based and self-training model, which shows effective domain transfer ability on actual clinical application and practice.

Related works

Domain shift (19), which refers to the performance degradation of deep learning models on input data from new, unseen domains, poses a critical impediment to the real-world clinical deployment. To address this issue, DA techniques have emerged as a promising family of solutions. These techniques can be broadly categorized into three paradigms based on the availability of labeled data in the target domain: supervised domain adaptation (SDA), which requires a fully labeled target dataset; semi-supervised domain adaptation (SSDA), which leverages a small set of labeled target data alongside a larger unlabeled target set; and UDA, which operates under the most challenging yet practical scenario by utilizing only unlabeled target data and labeled source data. Among these, UDA is more applicable to real-world clinical PD diagnosis settings, as it can be seamlessly applied to new medical center data without expert annotations. UDA concentrates on domain-invariant features learning that minimize the discrepancy between the source and target distributions, thereby enabling a model trained on labeled source data to generalize effectively to the target domain. Current UDA methodologies can be classified into discrepancy-based, adversarial-based, reconstruction-based and self-training methods. This categorization and the associated methodological evolution highlight UDA field’s progression from direct statistical alignment to more sophisticated, interactive, and robust frameworks that combine the strengths of multiple paradigms.

Discrepancy-based methods

Discrepancy-based approaches directly minimize a predefined statistical measure of distance between the feature distributions from the source and target domain, respectively. Common metrics include maximum mean discrepancy (MMD) (22), multiple kernel variant of MMD (MK-MMD) (23), correlation alignment (CORAL) (24), and margin disparity discrepancy (MDD) (25). Deep domain confusion (22) augments a standard classification network with an adaptation layer and an MMD loss term. It forces the network to learn representations that are both semantically meaningful and domain-invariant. Deep CORAL (24) advanced this concept by aligning not just the means but the covariances of the feature distributions, providing a more robust second-order statistical alignment. UDAS2S (26) applied patch comparison learning to constrain the consistency of local patches of source and target images for medical image segmentation. BFET (27) utilized bidirectional cross-attention transformer and prototype knowledge-based pseudo labeling for domain-invariant feature representations. Discrepancy-based methods offer the key advantages of training stability and a degree of interpretability. However, they primarily focus on aligning marginal distributions and may fail to capture complex, high-dimensional shifts. In addition, their effectiveness can also be sensitive to manual selection of an appropriate discrepancy metric and its associated weighting hyperparameters. It’s worth noting that discrepancy-based methods are often implemented alongside fine-tuning or adversarial training to align the feature distributions of both domains within a shared deep network.

Adversarial-based methods

Adversarial-based approaches are currently one of the most dominant and successful paradigms in UDA. These methods regard domain alignment as an adversarial game, inspired by generative adversarial networks (GANs) (28). Notable architectures in this family include the landmark domain-adversarial neural networks (DANN) (29), adversarial discriminative domain adaptation (ADDA) (30), and their numerous variants tailored for tasks such as segmentation and classification. DANN introduced a gradient reversal layer (GRL) to adversarially train a domain discriminator against the feature extractor, encouraging the extracted features indistinguishable between the source and target domains. ADDA decouples this process by first learning a source encoder and classifier, then learning a separate target encoder using a GAN-like loss to map target data into the source feature space. The conditional domain adversarial network (CDAN) (31) further refines this by conditioning the adversarial process on the classifier’s predictions, creating a more informative and powerful multi-linear conditioning for the discriminator. Then CDAN-SDAT (32) is proposed to smooth domain adversarial training for the training stability, and further combined with MCC loss (33) (CDAN-SDAT-MCC) for better performance. Further extending this line of work, MIC (34) combines CDAN-SDAT-MCC with masked target image prediction consistency objective to capture image context. Adversarial-based methods obtain advantages of powerful alignment capacity and wide applicability for both feature-level and pixel-level adaptation. However, they are also known for training instability, such as the mode collapse and oscillatory behavior common in GAN-based adversarial training and it is also sensitive to the careful tuning of hyperparameters.

Reconstruction-based methods

Reconstruction-based approaches leverage auxiliary reconstruction tasks to ensure the learned domain-invariant features preserve critical semantic content and avoid mode collapse. DRCN (35), a representative approach, utilizes a dual stream autoencoder to learn a shared representation that can both classify labeled source samples and reconstruct unlabeled target samples, ensuring the features retain semantic information critical for the task. PixelDA (36) employs a GAN to perform pixel-level domain translation. It translates source images into the style of the images of target domain, thereby creating a synthetic, labeled target-domain dataset for training a task-specific model. Despite their strengths, reconstruction-based methods face the challenge of task misalignment, where the objective of reconstructing input data may not directly enhance performance on the discriminative task of interest. Additionally, these methods require designing and training additional decoder networks, lead to the increase of architectural complexity and computational cost.

Self-training methods

Self-Training approaches leverage the model’s own predictions (pseudo-labels) on the dataset of target domain to iteratively the adaption, by selecting high-confidence predictions and using them as supervisory signals for the target data. It often incorporates entropy minimization principles to encourage confident, low-entropy predictions on target samples. AST (37) employs two distinct classifiers on a shared feature extractor. Target samples where the classifiers disagree are used for adversarial feature alignment, while samples with confident, agreeing predictions generate pseudo-labels for self-training. Cluster Alignment (38) explicitly enforces the cluster assumption by minimizing the entropy of target predictions and aligning the cluster centers of source and target features, leading to more accurate and robust pseudo-labels. CRST (39) employs a mean-teacher model to generate stable, consistent pseudo-labels and combines this with entropy minimization to sharpen predictions and achieve effective adaptation. Despite their effectiveness, self-training methods are faced with the error accumulation caused by incorrect pseudo-labels and are also sensitive to the selection of confidence thresholds.

In conclusion, discrepancy-based methods offer stability and interpretability but are often outperformed by more powerful adversarial methods, which in turn grapple with training instability. Reconstruction-based methods provide a robust solution for learning generalizable features but may introduce computational overhead without a guaranteed boost in task-specific accuracy. A prominent trend in current research is the development of hybrid approaches that aim to synergize the strengths of these categories while mitigating their weaknesses. For instance, combining adversarial alignment with self-training (e.g., AST) can leverage the strong feature alignment of the former and the target-specific learning of the latter. Zhu et al. also combined self-training and adversarial training into SSDA framework The MIC framework further improves this insight by enforcing the consistency between predictions of masked target images to learn spatial context relations of the target domain. MAPSeg (40) also unifies masked Autoencoding and pseudo labeling for cross-domain brain and cardiac image segmentation. However, it still faces the challenge of incorrect pseudo labels, as they only utilize the max prediction probabilities of target images as pseudo labels. Therefore, we introduce curriculum learning that dynamically adjusts the confidence threshold for generating pseudo-labels, thereby improving the quality of pseudo-labels and enhancing feature representation learning across both source and target domains. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0139/rc).

Methods

Due to acquiring expert annotations for every new clinical site is prohibitively expensive and time-consuming, UDA only requires labeled data from the source domain and unlabeled data from the target domain, which is suitable for multi-center PD diagnosis in real-world clinical settings. As shown in Figure 1, We propose curriculum-guided unified UDA (CGU-UDA) framework that combines adversarial-based methods and self-training methods and introduces dynamic threshold computation for accurate pseudo label generation. Meanwhile, inspired by MIC, we utilize masked target image prediction consistency to learn spatial context relation. CGU-UDA consists of three synergistic sub-modules that is adversarial UDA for domain-invariant feature learning, curriculum pseudo labeling (CPL) for adaptive class-wise thresholding and target masked image consistency training (TMICT) for context-aware feature regularization. CPL and TMICT are jointly optimized within an adversarial self-training loop, as illustrated in Figure 1. We chose 3D ResNet with 18 layers (3D ResNet-18) as the backbone, where 3D convolution replaces the 2D convolution in original ResNet. Source MRI images are input to 3D ResNet-18 for the supervised learning, meanwhile original and masked target MRI images are respectively input to the model for feature extraction. Then source and target features are input to a domain discriminator for discriminating the domain of the input features. This constructs an adversarial based UDA framework. TMICT is used for the consistent constraint between original and randomly masked target image through the pseudo label generated by CPL module. Target MRI images are input to Teacher network which is the exponential moving average (EMA) of the classification network (student network) to get pseudo labels through curriculum learning based dynamic threshold obtained by CPL module and the predictions of masked target MRI images are constrained consistently with the pseudo label for self-training. Teacher network produces initial prediction probability from target images, and its prediction confidence statistics are continuously monitored to estimate per-class learning effects. These estimates dynamically adjust the class-wise thresholds, which in turn govern which pseudo-labels are retained to supervise the student network on original and randomly masked target images. Through adversarial learning and masked target consistency constraint with dynamic threshold, our proposed algorithm obtains better feature adaptation and spatial context relation modeling.

Figure 1 Proposed CGU-UDA framework for multi-center PD diagnosis. CGU-UDA, curriculum-guided unified unsupervised domain adaptation; EMA, exponential moving average; MCC, minimum class confusion; MRI, magnetic resonance imaging; PD, Parkinson’s disease.

Data processing

We utilized MRI data from two independent cohorts for PD diagnosis. The first cohort comprised 234 subjects (119 PD patients and 115 healthy controls) obtained from the publicly available PPMI database (http://www.ppmi-info.org), acquired using a Siemens 3.0T MRI scanner. The second cohort consisted of 188 subjects (102 PD patients and 86 healthy controls) collected from Henan Provincial People’s Hospital using a Siemens 3.0T Prisma MRI scanner equipped with a 64-channel head-neck coil. Each data in the multi-cohort datasets is naturally separate with no data leakage. To ensure a rigorous evaluation and prevent data leakage, we performed a patient-level split for all cross-sectional data. All scans belonging to the same patient were assigned exclusively to either the training set or the test set. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Henan Provincial People’s Hospital [No. (2024) Ethics Review-196] and all patients provided written informed consent.

Fractional anisotropy (FA) maps derived from diffusion tensor imaging (DTI) were employed as input to the network. FA quantifies directional water diffusion and serves as a sensitive biomarker for assessing white matter integrity and detecting early axonal degeneration in PD. FA images are calculated by DTI preprocessing using the FMRIB Software Library (FSL 6.0; Oxford, UK; https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSL). Preprocessing included eddy current correction with gradient direction adjustment, brain extraction, and tensor fitting to derive FA imaging. The original FA images with a resolution of 94×94×67 were spatially normalized to MNI standard space and resampled to 91×109×91. To focus on salient region relevant to early PD pathology and reduce computational complexity, the resampled FA images were cropped to 91×109×71. This cropping removed only the top and bottom slices along the z-axis, while the retained region still covered the deep gray matter nuclei (substantia nigra, striatum, globus pallidus, and thalamus) and the corpus callosum most relevant to PD. The preprocessed FA images were subsequently used as model input.

Adversarial UDA framework

Adversarial UDA framework for MRI images is based on CDAN-SDAT with additional MCC loss, which is an adversarial-based UDA framework designed to learn transferable features by conditioning the domain adversarial adaptation process on the discriminative information from the classifier’s predictions. This conditioning enables more effective and discriminative alignment of multimodal distributions across domains. The implementation consists of three core components: a feature extractor G_f, a task classifier G_y, and a domain discriminator G_d.

Let xiS and xjT denote samples from the labeled source domain D_s and the unlabeled target domain D_t, respectively. The feature extractor G_f, parameterized by weights θ_f, maps an input MRI image to a deep feature representation, Fi=Gf(xi). The task classifier G_y with parameters θ_y then processes this representation to produce a class probability prediction. The model is trained on the source domain using the standard cross-entropy loss for the classification task:

Ls=LCE(Gy(Gf(xS)),yS)[1]

Where L_CE is the cross-entropy loss function.

The central innovation of CDAN is the design of its domain discriminator G_d. Instead of receiving only the feature vector as input, G_d is conditioned on the outer product of the feature representation and the classifier’s prediction, that is f⊗p. This multilinear map captures the complex, multiplicative interactions between the feature and its predicted class, providing a richer signal for discrimination. The domain discriminator, parameterized by θ_d, is trained to distinguish between source (domain label 1) and target (domain label 0) examples by minimizing the following binary cross-entropy loss:

Ld=−ExS∼Dslog[Gd(fS⊗pS)]−ExT∼Dtlog[1−Gd(fT⊗pT)][2]

Concurrently, the feature extractor G_fis adversarially trained to maximize this discriminator loss, thereby learning to generate features that confuse the discriminator. This is achieved by incorporating a GRL between the feature extractor and the domain discriminator. The GRL leaves the input unchanged during the forward pass but reverses the gradient by multiplying it by a negative factor (−λ) during the backward pass, where λ is a scheduling parameter that gradually increases from 0 to 1 during training.

The full objective function of CDAN is a minimax optimization problem that combines the classification loss and the adversarial domain loss:

minGf,GymaxGdλsLs+λdLd[3]

Where λ_s and λ_d are hyperparameters that balance the contribution of the DA term. Through this adversarial game conditioned on the classifier’s predictions, CDAN learns features that are not only domain-invariant but also tightly coupled with the decision boundaries of the task, leading to superior transfer performance on the target domain.

Masked target image consistency training

A pivotal component of our framework is the TMICT inspired by MIC and extended to 3D MRI image field, which is designed to explicitly enhance the model’s ability to leverage contextual relationships within the target domain. This module operates on unlabeled target MRI images and is seamlessly integrated into the overall training objective. The procedural flow for applying MIC to a target domain sample is illustrated in Figure 1 and elaborated in the following steps.

Given an unlabeled target MRI image xT∈ℝH×W×D×C, we first generate a binary patch mask M∈{0,1}H×W×D×C to selectively withhold local information. The mask is constructed by dividing the image into non-overlapping patches of size b×b×b. Each patch is independently masked with a probability r (the mask ratio). The resulting masked target image x^M is then computed via element-wise multiplication:

xM=M⊙xT[4]

Here, the masked regions are set to zero, effectively removing the corresponding local visual cues and compelling the network to rely on the surrounding contextual information for its predictions.

The masked image x^M is fed into the feature extractor G_f and task classifier G_y, which produces a softmax prediction y^M:

y^M=Gy[Gf(xM)][5]

Due to the absence of visual content in the masked regions, the feature extractor is forced to infer the semantics of these areas based solely on the context provided by the unmasked portions of the image.

To provide a supervisory signal for the masked prediction in the absence of ground truth labels, we employ a self-training strategy using pseudo-labels. A teacher network g_∅, which is an EMA of the feature extractor and task classifier (student network) weights, is employed for this purpose. The teacher’s weights ∅ are not trained by gradient descent but are updated at each training step t from the student’s weights θ using a smoothing factor α:

ϕt+1←αϕt+(1−α)θt[6]

Where t denotes the training step. This teacher network processes the original, uncorrupted target MRI image x^T, benefiting from both full local appearance and global context, to generate a high-quality classification prediction. The one-hot pseudo-label y^T is then obtained by taking the argmax of the teacher’s output probability distribution p^T.

The loss is designed to enforce consistency between the student’s prediction y^M from the masked image and the teacher-generated pseudo-label y^T. This is achieved by minimizing a cross-entropy loss between the two distributions. To mitigate the negative impact of potentially erroneous pseudo-labels, especially in the early training stages, we weigh the loss for each MRI image using a quality estimate q^T, which reflects the confidence of the teacher’s prediction. q^T is defined as the predicted softmax probability of pseudo-label.

The final TMICT loss L_M for a single target MRI image is thus computed as:

LM=qT⋅LCE(y^M,yT)[7]

The TMICT module is integrated as an additional unsupervised loss term into the above-mentioned adversarial UDA optimization framework. The total objective function combines the supervised loss on the labeled source domain and the unsupervised losses on the target domain:

minθλsLs+λMLM+λdLd[8]

The hyperparameters λ_s, λ_M and λ_d control the balance between the different objectives. Through this optimization, the student network progressively learns to utilize comprehensive contextual relationships in the target domain for accurate semantic prediction, thereby achieving more robust DA. Over multiple training iterations, as different random patches are masked, the model learns comprehensive contextual relationships present in the target domain, leading to more robust performance, particularly for classes with ambiguous local appearance.

CPL

The TMICT module relies on a fixed, pre-defined confidence threshold τ for all classes throughout training and the final performance is easily influenced by this predefined threshold. This approach fails to account for the varying learning difficulties and progress across different classes. To address this, we introduce a curriculum learning based pseudo labeling strategy that dynamically adjusts the confidence threshold for each class based on its estimated learning status. This allows the model to progressively utilize more unlabeled data from classes that are harder to learn. The core of CPL is to replace the fixed threshold τ with a class- and time-specific flexible threshold τt(c) for class c at training step t.

First, estimate the class-wise learning effect. The learning status of a class is estimated by the number of unlabeled samples whose predicted confidence for that class exceeds the fixed base threshold τ. Formally, for class c at step t, this quality σt(c) is defined as:

σt(c)=∑n=1N1condition[max(ptT>τ)]1condition[argmax(ptT)=c][9]

Where N is the total number of unlabeled samples, ptT is the model’s predicted probability distribution for target sample at step t. A higher σt(c) indicates that the model can confidently predict more samples as belonging to class c, suggesting a better learning effect for that class.

To facilitate comparison across classes, the raw count σt(c) is normalized relative to the maximum count among all classes at step t.

βt(c)=σt(c)maxc′σt(c′)[10]

This normalization ensures that βt(c)∈[0,1], with the best-learned class having βt(c)=1. Then βt(c)=1 is used to scale the fixed threshold τ to obtain the dynamic threshold τt(c).

τt(c)=βt(c)τ[11]

The resulting flexible threshold τt(c) is then used in the unsupervised loss calculation. For the consistency computation between masked target MRI image and target MRI image, it contributes to the loss only if its maximum confidence score exceeds the threshold specific to its predicted class τt[argmax(ptT)]. Above-mentioned TMICT loss LM in CPL-enhanced methods can be formulated as:

LM=1N∑n=1N1condition[max(ptT>τ(c))⋅LCE(y^M,yT)][12]

This curriculum-based thresholding strategy encourages the model to focus on less-learned classes by lowering their entry barriers (thresholds) early on, while progressively demanding higher confidence for well-learned classes. The process is computationally efficient, requiring only minimal extra bookkeeping [σt(c) counts] during the standard forward pass for unlabeled data, with no additional backward or forward propagations.

Results

Parameters setting

The training parameters of proposed CGU-UDA network is similar to common CDAN-SDAT network. We trained for 30 epochs using the SGD optimizer with an initial learning rate of 2×10⁻³, weight decay of 1×10⁻³ and momentum of 0.9. The smoothness parameter is set to 0.02 for SDAT adversarial training. The masked block size is set to 32×32×32 with a mask ratio of 0.7. The model was implemented in PyTorch framework and trained on an NVIDIA A100 GPU. For model performance evaluation, we use accuracy, precision, recall, F1 score, and the area under the curve (AUC) metrics.

Performance comparison with UDA methods

CGU-UDA network combines adversarial training using domain discriminator and CPL for masked target MRI image consistency training for multi-center PD diagnosis. This method learns the spatial context and transferable features. For better verifying the performance of proposed CGU-UDA network, we compare it with previous adversarial training UDA methods as shown in Table 1. The performance of 3D ResNet-18 is obtained by directly training on one domain data and testing on the other domain data. It is obvious that proposed CGU-UDA network achieves the best classification performance not only on Hospital→PPMI adaptation task, but also on PPMI→Hospital adaptation task. DALN (41) is discriminator-free adversarial learning network whose category classifier is reused as a discriminator to address the mode collapse problem. However, it obtained poorer performance than other adversarial training UDA network with discriminator. Most adversarial training UDA algorithms achieve better performance compared to original 3D ResNet-18 and DALN which proves the advantage of discriminator-based adversarial training UDA framework. It also shows a clear tendency that the performance becomes better with condition, SDAT, MCC loss, masked consistency and adaptive pseudo labeling module adding, especially on Hospital→PPMI adaptation task. With proposed curriculum learning based pseudo labeling adding, CGU-UDA network obtained over 2% accuracy and over 6% AUC value improvement on Hospital→PPMI adaptation task, and over 2% accuracy and over 3% AUC value improvement on PPMI→Hospital adaptation task compared to MIC.

Confusion matrix and ROC curve comparisons of MIC and CGU-UDA are shown in Figure 2. It is obvious that CGU-UDA achieves the best AUC value, and the most accurate prediction of healthy and PD case on PPMI→Hospital adaptation task with 57 healthy and 67 PD correct predictions. CGU-UDA obtained better healthy case diagnosis accuracy (true negative ratio). Although CGU-UDA yielded a slightly lower PD diagnosis accuracy (true positive ratio) than MIC on the Hospital→PPMI adaptation task with only 4 wrong PD predictions, it achieved a substantially better healthy case diagnosis accuracy (true negative ratio) and lower false positive ratio. From a clinical perspective, minimizing false positives is critical in PD diagnosis because classifying a healthy individual as PD can cause unnecessary patient anxiety, lead to extra diagnostic tests, and result in unnecessary use of dopaminergic medication and its side effects. Thus, the lower false positive rate of CGU-UDA helps reduce these risks and offers a more favorable risk-benefit balance in real-world multi-cohort PD diagnosis.

Figure 2 Confusion matrix and ROC curve comparisons. MIC and CGU-UDA are compared on the two different directions including Hospital→PPMI and PPMI→Hospital through confusion matrix and ROC curve. CGU-UDA, curriculum-guided unified unsupervised domain adaptation; MIC, masked image consistency; PPMI, Parkinson’s Progression Markers Initiative; ROC, receiver operating characteristic.

We also compare proposed CGU-UDA network with other discrepancy-based (DAN, MMD, DeepCoral and MDD) and adversarial-based (ADDA) methods for better reliable comparison. As shown in Table 2, proposed CGU-UDA network achieves the best performance on five evaluation metrics among these models. CGU-UDA network gets a large performance gain on Hospital→PPMI task, with maximum over 20% accuracy improvement and 23% AUC improvement, and minimal 17% accuracy improvement and 10% AUC improvement. The improvement of CGU-UDA network is relatively less on PPMI→Hospital task, with only maximum over 4% accuracy increase and 3% AUC increase. The performance of these algorithms is completely different on these two adaptation tasks that model often performs poor on Hospital→PPMI task and a little better on PPMI→Hospital task, which demonstrate the instability of these algorithms and are easily influenced by dataset. Compared to adversarial based ADDA algorithm, CGU-UDA network utilizes masked target MRI image prediction consistency constraint with dynamic threshold for more accurate pseudo label generation to learn image context for better feature learning and therefore obtains better performance. Compared to discrepancy-based algorithms, CGU-UDA network does not exploit the discrepancy computation between the source and target MRI images but utilizes the target own consistency and source-target feature adversarial training to capture better and domain invariable features.

Ablation experiments

CGU-UDA network utilizes consistency constraint between original and masked target MRI image for spatial context learning and applied random 3D binary masks to multiply original images for masked target MRI image generation. This is an input mask manner. For evaluating the effectiveness of simple random mask on original input, we compare it with feature mask achieved by dropout and DropBlock (42). Dropout is a regularization method that randomly drops some neurons with a predefined probability and improves the model generalization performance. Therefore, dropout can be seen as a feature mask manner, and we applied dropout into the extracted high-level features by the final residual block of 3D ResNet-18. Additionally, we applied DropBlock module into the extracted low-level features by the first 3D convolution of 3D ResNet-18 for evaluation. DropBlock discards a consecutive adjacent region of the feature map together which is similar to random input mask manner. The drop probability of dropout and DropBlock is set to 0.5 and the block size of DropBlock is set to 15 to maintain mask size consistency with CGU-UDA network (the down-sample rate of first convolution is 2). The comparison results are shown in Table 3. No masked network means the original CDAN-MCC-SDAT network.

Results demonstrate that CGU-UDA network obtain the best performance compare to no masked and other masked manners on Hospital→PPMI adaptation task, but a little weaker performance compare to DropBlock masked manner on PPMI→Hospital adaptation task. CGU-UDA network is always superior to no masked and Dropout masked manners on two adaptation tasks. High-level features dropout masked network always gets better performance compared to the no masked with max 5% accuracy improvement on PPMI→Hospital adaptation task. However, DropBlock masked manner obtains unstable performance on these two transfer tasks, which gets a most poor performance on Hospital→PPMI adaptation task with over 8% accuracy and AUC decrease and best performance on PPMI→Hospital adaptation task with almost 6% accuracy and AUC increase compared to no masked manner. These comparisons reveal the advantage of targe MRI image masked consistency constraint which better capture the image context for feature learning. The suboptimal performance of Dropout masked manner might be attributed to the high-level feature mask with sufficient semantic information which is difficult for reconstruction and further influences the final performance. Meanwhile the DropBlock masked manner is influenced by the feature representation ability of first convolution, thus achieves different performance presentation on different adaptation tasks. Our proposed simple and direct masked manner on original MRI images can avoid the influence of network capacity and difficulty of high-level feature context learning, and better capture the primordial image information and context, thus can achieve the best performance.

When constraining the consistency between original and masked target MRI image, CGU-UDA network utilizes curriculum learning based pseudo labeling to achieve a dynamic threshold for accurate pseudo label generation. MIC directly applies max probability for pseudo label generation and consistent loss computation, and most semi-supervised algorithms use fixed threshold by hyperparameter setting to generate pseudo label. There are also some adaptive threshold algorithms to improve the confidence of pseudo label, such as ADSH (43) deployed to consider the number of pseudo labels selected for each class and minimize empirical risk. Therefore, we applied these threshold methods into CGU-UDA network framework for pseudo labeling evaluation. Fixed threshold is set to 0.9 in our experiment.

As shown in Table 4, curriculum learning based pseudo labeling achieves the best performance compared to other threshold algorithms. Compared to fixed threshold, CGU-UDA network obtains an accuracy improvement of over 3% and an AUC improvement of nearly 10% on Hospital→PPMI, and an accuracy improvement of over 3% and an AUC improvement of over 2% on PPMI→Hospital. However, the performance increase compared to ADSH is relatively less, with only 0.9% accuracy increase on Hospital→PPMI and 0.5% accuracy increase on PPMI→Hospital. Meanwhile ADSH also get a better performance compared to MIC and fix threshold for pseudo labeling. The reason is that these two threshold methods all deployed a dynamic threshold algorithm either according to the learning state or selected pseudo labels counts of each class. Fixed threshold achieves a worse performance compared to max probability on the Hospital→PPMI task, and yields a slightly higher AUC compared to max probability on the PPMI→Hospital task. Fixed thresholding can occasionally maintain competitive ranking performance on less challenging adaptation task. MIC with max probability explicitly uses the pseudo label generated by max probability to compute consistent loss with a max probability weight, while fixed threshold only uses the pseudo label generated by the sample that the predicted max probability exceeds the threshold. The reduced and wrong sample consistent loss might decrease the learning effect due to the unselected loss computation, further weakens feature learning ability of the whole model, especially on the challenging adaptation task. ADSH and CGU-UDA all deploy a dynamic threshold that can get more and more accurate pseudo label and consistent loss constraints, thereby acquire better model classification performance.

To further assess the impact of class-wise curriculum learning, we compared the CPL with an entropy-based sample uncertainty estimate. We evaluated three variants that is Entropy, Entropy + CPL and Entropy × CPL. Entropy means that threshold is adjusted solely based on sample-wise entropy of the teacher’s softmax output without the class-wise CPL. Entropy + CPL means that class-wise CPL and entropy are both used and selected samples are the sum of the sample satisfying each criterion. Entropy × CPL means that class-wise CPL and entropy are both used and samples are selected only when they satisfy both criteria. The quantitative results on the two cross‑domain tasks are shown in Figure 3.

Figure 3 Performance comparisons between entropy uncertainty and curriculum learning threshold. AUC, area under the curve; CPL, curriculum pseudo labeling; PPMI, Parkinson’s Progression Markers Initiative.

Entropy yields moderate performance in both tasks, but it is inferior to the CPL approach on Hospital→PPMI task and superior on PPMI→Hospital task. This means that class-wise learning effects provide valuable information beyond sample-wise confidence on hard domain adaption task and sample-wise confidence could achieve better performance on simple task. Conversely, Entropy + CPL leads to severe performance degradation on both Hospital→PPMI and Hospital→PPMI tasks, CPL alone remains the better performance, indicating that class-level learning statistics are more reliable when the source model is less confident overall. The combination with entropy on difficult Hospital→PPMI task might introduce uncertain prediction which influences the consistency constraint learning with wrong pseudo label. We hypothesize that the multiplicative combination allows the entropy to serve as a continuous modulator of the threshold, rather than a hard gate, preserving the curriculum’s ability to adapt to diversified samples while down-weighting potentially misclassified samples with high uncertainty. On PPMI→Hospital, multiplicative combination (Entropy × CPL) yields the best overall performance, particularly improving AUC by 3.89 percentage points over CPL alone. But it achieves a lower performance on Hospital→PPMI task. This suggests that strict filtering that samples must simultaneously satisfy both a class-wise dynamic threshold and an entropy-based confidence criterion is beneficial and can select more confident samples for consistency constraint, but reversely overly strictness easily exclude more effective information and severely limit the supervisory signal for the more challenging domain shift for Hospital→PPMI task. This suggests that in the more challenging adaptation direction, both criteria being satisfied simultaneously is overly restrictive.

The results demonstrate that while sample-wise uncertainty provides complementary information, its integration must be carefully designed to avoid excessive pruning of the training signal. The multiplicative integration is more effective, as it preserves the curriculum’s adaptability while incorporating instance-level confidence. The optimal integration strategy appears task-dependent, likely influenced by the magnitude of the domain gap and the reliability of the source-initialized model. We adopt CPL as the default setting for our framework, as it offers the best balance across both adaptation directions.

We also investigated the impact of the initial confidence threshold τ on the performance of our curriculum learning guided pseudo-labeling mechanism. The threshold governs the base confidence level required for a pseudo-label to be considered, with class-wise dynamic adjustments. Experiments used threshold values τ∈{0.8, 0.85, 0.9, 0.95, 0.99}. The results are summarized in Figure 4.

Figure 4 Performance effect with different initial thresholds of curriculum learning. AUC, area under the curve; PPMI, Parkinson’s Progression Markers Initiative.

For the Hospital→PPMI task, the optimal initial threshold is 0.95, achieving the highest values across all metrics. Notably, a threshold of 0.80 yields moderate performance, while thresholds of 0.85 and 0.90 result in lower accuracy. The sharp drop at 0.99 suggests that an overly stringent threshold excludes too many informative samples, especially in the early training phase, thereby limiting the supervisory signal for adaptation. Conversely, in the PPMI→Hospital task, the performance is more balanced. The best accuracy (67.02%) and F1 (66.96%) are obtained at τ=0.85, and the second accuracy, precision, recall and F1 are obtained at τ=0.95. While the highest AUC (68.84%) is observed at τ=0.99, indicating that a higher threshold benefits the ranking ability in this direction.

Comparing the two tasks, we observe that the optimal threshold is domain‑direction dependent. Hospital→PPMI benefits from a higher initial threshold (0.95), possibly because the source domain (Hospital) data are more homogeneous and the model quickly gains confidence, whereas PPMI→Hospital exhibits a broader optimal range (0.85–0.99) with less sensitivity. A plausible explanation is the difference in data distribution complexity that the PPMI dataset with multi‑center data may contain more inherent variability, making the curriculum learning less sensitive to the exact initial threshold. Based on these results, we adopted τ=0.95 for all tasks in our main experiments, which consistently yielded strong results. The curriculum learning mechanism effectively compensates for suboptimal initial choices to some extent, as evidenced by the non‑monotonic relationship between τ and final performance. However, extreme values (e.g., 0.99 in the Hospital→PPMI task) should be avoided as model could not achieve higher confidence on another domain due to domain difference which result in limited learning of masked consistency.

For further evaluating the importance and impact of the pseudo label on model performance, we adopt a retraining manner that adding the top-5 confident pseudo label into the source MRI training images with every five epochs for model training (named as Add-retraining). Results in Table 5 show that Add-retraining obtains a poor performance compare to CGU-UDA with no Add-retraining on Hospital→PPMI and PPMI→Hospital adaptation task, although the performance decrease is little on Hospital→PPMI adaptation task with only 0.43% accuracy decrease. Add-retraining obtains over 3% accuracy decrease and 7% AUC decrease on PPMI→Hospital adaptation task. This demonstrates the crucial impact of pseudo label accuracy for model performance again, potentially verifies the importance of threshold for pseudo label, and provides experimental evidence for pseudo labeling method improvement.

We further investigated the sensitivity of our framework to the mask ratio, a critical hyperparameter in the masked consistency module. As shown in Table 6 and Figure 5, the optimal mask ratio is task-dependent. For the Hospital→PPMI task, performance peaks sharply at 0.7 across all metrics (accuracy: 67.95%, AUC: 0.7148), while lower (0.3, 0.5) and higher (0.9) ratios all result in substantial performance degradation, with AUC dropping to 0.4900 at 0.5 and 0.5315 at 0.9. This non-monotonic behavior suggests that the model benefits from a sufficiently challenging masking ratio that encourages robust contextual learning, but excessive corruption (0.9) removes too much structural information. In contrast, PPMI→Hospital shows relatively stable performance across different mask ratios, with a slight preference for 0.3 with an accuracy of 67.55% and AUC of 0.7110. The differences across ratios are relatively small with an accuracy range of 65.96–67.55%, suggesting that this adaptation direction is less sensitive to the mask ratio. This difference may reflect the greater inherent heterogeneity of the multi-center PPMI data, which makes the model less sensitive to the specific masking configuration. Based on these results, we adopt a mask ratio of 0.7 in our main experiments.

Figure 5 Accuracy and AUC comparisons under different mask ratios. AUC, area under the curve; PPMI, Parkinson’s Progression Markers Initiative.

Explainability analysis

To assess how well the adapted features align with established PD neuropathology, we compared the class activation heatmaps of CGU-UDA and the strongest baseline, MIC in Figure 6. In the Hospital→PPMI direction, both methods activated the corpus callosum, but MIC further diffused its attention into the frontal and parietal cortex, whereas CGU-UDA remained more concentrated on the corpus callosum. In the more challenging PPMI→Hospital direction, MIC largely departed from the corpus callosum and activated mainly the frontal and parietal cortex, while CGU-UDA consistently localized to the corpus callosum across subjects. The corpus callosum is one of the principal regions affected by early white matter degeneration in PD, where reduced FA has been repeatedly associated with motor and cognitive impairment (44), whereas frontal and parietal cortical changes are not characteristic pathological features of PD. Compared with the strongest baseline MIC, the attention learned by our CGU-UDA is therefore more consistent with previously reported PD white matter degeneration patterns, and this advantage is particularly pronounced in the more difficult adaptation direction, in line with the quantitative trends in Table 1. It should be noted that, beyond the target region, the heatmaps also include some non-relevant areas. This is because Grad-CAM++ weights the low-resolution feature maps of the last convolutional layer and upsamples them to the input size, producing blob-like saliency with blurred boundaries that cannot be precisely matched to a single anatomical structure. This overlap with neighboring tissue reflects the limited resolution of the attribution method rather than the actual basis of the model’s discrimination. The activation peaks are stably located along the midline callosal course within the brain parenchyma and recur consistently across subjects and both adaptation directions, indicating that the model relies on stable, anatomically meaningful features.

Figure 6 Class activation heatmaps comparisons on PD subjects. (A) Heatmaps of MIC on Hospital→PPMI; (B) heatmaps of CGU-UDA on Hospital→PPMI; (C) heatmaps of MIC on PPMI→Hospital; (D) heatmaps of CGU-UDA on PPMI→Hospital. PD, Parkinson’s disease.

Discussion

The early and accurate diagnosis of PD is a critical challenge in clinical neurology, directly impacting patient management and therapeutic outcomes. MRI serves as a vital tool for detecting subtle neurodegenerative changes. However, the development of robust, clinically deployable AI models is significantly hampered by the heterogeneity of MRI data that collecting from diverse scanners and protocols across institutions. UDA techniques are applied to solve this domain shift problem. While UDA approaches that combines self-training with adversarial learning have advanced feature discriminability, but face with error propagation and confirmation bias caused by noisy target domain predictions. To tackle this core limitation, we proposed CGU-UDA framework that innovatively integrates a curriculum learning mechanism to dynamically adjust confidence thresholds for pseudo-label selection based on real-time model performance, thereby progressively refining label quality. Furthermore, we introduced a masked image consistency constraint to force the model to learn robust, context-aware features from partially masked target data.

The proposed CGU-UDA framework offers distinct and significant advantages over existing UDA methods. These advantages can be categorized into the mitigation of confirmation bias, the enhancement of contextual feature learning and generalizability to other MRI modalities.

Mitigation of confirmation bias via adaptive thresholding

A fundamental weakness of standard self-training UDA methods, including FixMatch-inspired approaches, is their reliance on a fixed confidence threshold for pseudo-label selection. This static strategy is predicated on an unrealistic assumption of uniform learning progress across all classes and feature spaces. Actually, only a few easy or biased samples may surpass a high threshold leading to sparse and potentially unrepresentative supervision in the early adaptation training. Conversely, a low threshold brings excessive noise. Both scenarios can increase confirmation bias and make model learn from its own initial errors. CGU-UDA introduces a curriculum-based dynamic threshold and decrease this bias. Our threshold is directly governed by the estimated learning effect, which quantifies how many samples per class have been reliably learned. This creates a self-correcting feedback loop: for poorly learned classes, the threshold is automatically lowered to gather more supervisory signal and accelerate their learning; for well-learned classes, the threshold raises to filter out ambiguous samples and refine feature specificity.

Enhanced contextual robustness through targeted consistency regularization

While adversarial training promotes global feature invariance, it can be insufficient for learning locally ambiguous yet context-dependent patterns in medical imaging (e.g., distinguishing adjacent anatomical structures). Recent methods like MIC have demonstrated the power of masked consistency for context learning. CGU-UDA innovatively integrates this concept within an adversarial self-training framework. Unlike MIC, which applies consistency between masked images and a static teacher’s pseudo-labels, our framework enforces consistency using the dynamically refined pseudo-labels from our own curriculum mechanism. This creates a tighter coupling: the adversarial network learns features that must be consistent under masking, while the masking task itself is supervised by increasingly accurate pseudo-labels. Compared to adversarial-only methods or consistency-only methods, this dual-regularization strategy forces the model to extract features that are simultaneously domain-invariant (via adversarial loss) and contextually consistent (via masked consistency), yielding representations that are more generalizable and anatomically plausible.

Generalizability to other MRI modalities

FA derived from DTI has been widely used as a sensitive marker for detecting white matter changes in PD. Reduced FA values in the substantia nigra and nigrostriatal pathway have been consistently reported in PD patients compared to healthy controls, reflecting the loss of dopaminergic neurons and the disruption of related fiber tracts. In addition, decreased FA in the corpus callosum and major association tracts have been associated with motor symptoms and cognitive decline in PD (44). These findings suggest that FA can serve as a useful imaging indicator for the early diagnosis and progression monitoring of PD, therefore we utilize FA as the research data. Although FA images were employed as a sensitive indicator of white matter integrity in PD, the proposed CGU-UDA framework is inherently modality-agnostic, as it operates on feature representations extracted from any input image, regardless of the MRI contrast (e.g., FA, T1-weighted, T2-weighted). The core innovations of CGU-UDA are curriculum-guided dynamic thresholding and masked consistency regularization which do not rely on any modality-specific properties. Therefore, the framework can be directly applied to T1-weighted imaging or other structural and diffusion MRI sequences without architectural modifications. CGU-UDA is readily transferable to other commonly acquired modalities such as T1-weighted MRI, which may further enhance the model’s clinical utility and generalizability.

Future work

The proposed CGU-UDA framework also shows strong potential for improving diagnostic generalizability in real-world clinical settings by directly addressing the core technical bottleneck of AI generalizability across diverse clinical settings. The most immediate application lies in deploying CGU-UDA models as a decision-support system within hospital radiology departments. A primary obstacle to current AI models is their “brittleness” when faced with data from a new scanner, even from the same manufacturer. CGU-UDA’s core strength is mitigating this, allowing a single model to maintain high diagnostic accuracy across the heterogeneous MRI equipment. This capability is crucial for ensuring equitable diagnostic quality, preventing performance drops in different medical institutions. Additionally, CGU-UDA is not only evaluated on cross-sectional MRI scans (Hospital data) but also on longitudinal data (PPMI data) that captures disease progression over time. To maintain the authenticity of results, longitudinal data of the same people only existed in train or test dataset. This reveals that proposed CGU-UDA model can effectively adapt across time points or detect disease conversion, capture temporal changes and facilitate early detection of disease conversion or progression applications.

Despite its promising results, there still exist several questions of the CGU-UDA framework for future work. Firstly, the current curriculum learning strategy for dynamic thresholding, while effective, primarily operates on a class-wise basis. It may not fully account for instance-level variations in prediction confidence or the complex, fine-grained anatomical sub-structures within brain MRI relevant to PD. Future research could explore more fine-grained, spatial-aware curriculum scheduling or incorporate uncertainty quantification to better weight each pseudo-label. Secondly, the performance of the masked consistency module may be sensitive to the choice of masking strategy (e.g., patch size, mask ratio). Although we followed established heuristics, systematically optimizing these hyperparameters for neuroimaging data or developing adaptive masking mechanisms could yield additional gains. Thirdly, while our study focused on multi-cohort adaptation, the cohort sizes used in this study (234 and 188 subjects) are relatively small, which may limit the statistical power and generalizability of the findings. And clinical deployment faces broader challenges including out-of-distribution samples from completely unseen protocols or rare pathological presentations. Extending CGU-UDA to more open-set and larger multi-center datasets with diverse demographic and imaging characteristics or test-time adaptation scenarios would enhance its practical robustness. Finally, the translational impact of this work would be strengthened by prospective validation in a clinical workflow and exploration of the model’s interpretability. Linking the adapted features to known PD biomarkers could build greater trust among clinicians. In conclusion, to address the critical issue of noisy pseudo-labels, CGU-UDA framework provides a feasible thought for clinically adaptive AI diagnostic tools, and future efforts could be aimed at refining its adaptability, robustness, and interpretability.

Conclusions

The clinical deployment of early PD diagnosis models is often hindered by significant performance degradation when applied across different medical centers, primarily due to domain shifts in multi-source MRI data. To address this challenge, this paper proposes a curriculum learning-guided self-training adversarial unsupervised domain adaptation network (CGU-UDA). This framework effectively integrates adversarial training with self-training strategies and introduces a dynamic thresholding mechanism to generate accurate pseudo-labels for the target domain. Extensive evaluations on multi-cohort PD MRI datasets demonstrate that the proposed CGU-UDA outperforms existing state-of-the-art methods in the cross-domain classification task. Ablation studies further confirm the contribution of the masked image consistency module and the dynamic pseudo-labeling module to the model’s strong DA capability. These results underscore the potential of CGU-UDA for robust clinical application in real-world heterogeneous imaging environments. More importantly, our framework achieves a notably lower false-positive rate on the challenging PPMI→Hospital task. Reducing false positives is clinically critical which helps avoid unnecessary patient anxiety, prevents inappropriate dopaminergic medication with its potential side effects, and reduces the need for costly confirmatory testing. Future work will focus on more accurate pseudo label generation algorithms and the alignment issues with more centers of data, to promote the model generalization and clinical application.

Acknowledgments

We thank all the participants for their cooperation in this study.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0139/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0139/dss

Funding: This work was supported by the Key Research and Development Program of Henan Province (No. 251111312900), the Science and Technology Project of Henan Province (No. 242102311029), the Central Plains Science and Technology Innovation Leading Talent Project (No. 244200510015), and the Medical Science and Technology Research Project of Henan Province (No. SBGJ202303007).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0139/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Henan Provincial People’s Hospital [No. (2024) Ethics Review-196] and all patients provided written informed consent.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Ball N, Teo WP, Chandra S, Chapman J. Parkinson’s Disease and the Environment. Front Neurol 2019;10:218. [Crossref] [PubMed]
2. Tolosa E, Garrido A, Scholz SW, Poewe W. Challenges in the diagnosis of Parkinson’s disease. Lancet Neurol 2021;20:385-97. [Crossref] [PubMed]
3. Haruyama T, Zin TT, Sakai K, Mochizuki H, editors. A Study on Diagnosis of Parkinson’s Disease by Walking Video. 2021 IEEE 10th Global Conference on Consumer Electronics (GCCE); 2021: IEEE.
4. Virameteekul S, Revesz T, Jaunmuktane Z, Warner TT, De Pablo-Fernández E. Clinical Diagnostic Accuracy of Parkinson’s Disease: Where Do We Stand?. Mov Disord 2023;38:558-66. [Crossref] [PubMed]
5. Rizzo G, Copetti M, Arcuti S, Martino D, Fontana A, Logroscino G. Accuracy of clinical diagnosis of Parkinson disease: A systematic review and meta-analysis. Neurology 2016;86:566-76. [Crossref] [PubMed]
6. Beach TG, Adler CH. Importance of low diagnostic Accuracy for early Parkinson’s disease. Mov Disord 2018;33:1551-4. [Crossref] [PubMed]
7. Heim B, Krismer F, De Marzi R, Seppi K. Magnetic resonance imaging for the diagnosis of Parkinson’s disease. J Neural Transm (Vienna) 2017;124:915-64. [Crossref] [PubMed]
8. Camacho M, Wilms M, Mouches P, Almgren H, Souza R, Camicioli R, Ismail Z, Monchi O, Forkert ND. Explainable classification of Parkinson’s disease using deep learning trained on a large multi-center database of T1-weighted MRI datasets. Neuroimage Clin 2023;38:103405. [Crossref] [PubMed]
9. Zhang D, Zhou L, Yao J, Shi Y, He H, Wei H, Tong Q, Liu J, Wu T. Increased Free Water in the Putamen in Idiopathic REM Sleep Behavior Disorder. Mov Disord 2023;38:1645-54. [Crossref] [PubMed]
10. Dixit S, Bohre K, Singh Y, Himeur Y, Mansoor W, Atalla S, Srinivasan K. A comprehensive review on AI-enabled models for Parkinson’s disease diagnosis. Electronics 2023;12:783.
11. Esmaeilzadeh S, Yang Y, Adeli E. End-to-end Parkinson disease diagnosis using brain MR-images by 3D-CNN. arXiv preprint arXiv:180605233 2018.
12. Camacho M, Wilms M, Almgren H, Amador K, Camicioli R, Ismail Z, Monchi O, Forkert NDAlzheimer’s Disease Neuroimaging Initiative. Exploiting macro- and micro-structural brain changes for improved Parkinson’s disease classification from MRI data. NPJ Parkinsons Dis 2024;10:43. [Crossref] [PubMed]
13. Yang M, Huang X, Huang L, Cai G. Diagnosis of Parkinson’s disease based on 3D ResNet: The frontal lobe is crucial. Biomedical Signal Processing and Control 2023;85:104904.
14. Yagis E, De Herrera AGS, Citi L, editors. Generalization performance of deep learning models in neurodegenerative disease classification. 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM); 2019: IEEE.
15. Zhao H, Tsai CC, Zhu C, Zhou M, Wang JJ, Liu Y, editors. Deep Learning-Based Regional Sub-models Integration for Parkinson’s Disease Diagnosis Using Diffusion Tensor Imaging. International Conference on Image and Graphics; 2021: Springer.
16. Kaur S, Aggarwal H, Rani R. Diagnosis of Parkinson’s disease using deep CNN with transfer learning and data augmentation. Multimed Tools Appl 2021;80:10113-39.
17. Chen Q, Hong Y, editors. Smart: Self-weighted multimodal fusion for diagnostics of neurodegenerative disorders. Proceedings of the 32nd ACM International Conference on Multimedia; 2024.
18. Cui X, Xu Y, Lou Y, Sheng Q, Cai M, Zhuang L, Sheng G, Yang J, Liu J, Feng Y. Diagnosis of Parkinson’s disease based on feature fusion on T2 MRI images. International journal of intelligent systems 2022;37:11362-81.
19. Kondrateva E, Pominova M, Popova E, Sharaev M, Bernstein A, Burnaev E, editors. Domain shift in computer vision models for MRI data analysis: an overview. Thirteenth International Conference on Machine Vision; 2021: SPIE.
20. Liu X, Yoo C, Xing F, Oh H, Fakhri GE, Kang J-W, Woo J. Deep unsupervised domain adaptation: A review of recent advances and perspectives. arXiv preprint arXiv:220807422 2022.
21. Zhang B, Wang Y, Hou W, Wu H, Wang J, Okumura M, Shinozaki T. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems 2021;34:18408-19.
22. Tzeng E, Hoffman J, Zhang N, Saenko K, Darrell T. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:14123474 2014.
23. Long M, Cao Y, Wang J, Jordan M, editors. Learning transferable features with deep adaptation networks. International Conference on Machine Learning; 2015: PMLR.
24. Sun B, Saenko K, editors. Deep coral: Correlation alignment for deep domain adaptation. European Conference on Computer Vision; 2016: Springer.
25. Zhang Y, Liu T, Long M, Jordan M, editors. Bridging theory and algorithm for domain adaptation. International Conference on Machine Learning; 2019: PMLR.
26. Xiao Y, Chen C, Wang L, Yu J, Chen B, Fu X, Lu M. A novel semi-supervised domain adaptive method for cross-modality medical image segmentation. Vis Comput 2025;41:9139-54.
27. Hao Z, Wang S, Long S, Li Y, Chai H. Bidirectional feature enhancement transformer for unsupervised domain adaptation. Vis Comput 2024;40:6261-77.
28. Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, Courville A, Bengio Y. Generative adversarial networks. Communications of the ACM 2020;63:139-44.
29. Ajakan H, Germain P, Larochelle H, Laviolette F, Marchand M. Domain-adversarial neural networks. arXiv preprint arXiv:14124446 2014.
30. Tzeng E, Hoffman J, Saenko K, Darrell T, editors. Adversarial discriminative domain adaptation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; 2017.
31. Long M, Cao Z, Wang J, Jordan MI. Conditional adversarial domain adaptation. Advances in Neural Information Processing Systems 2018;31.
32. Rangwani H, Aithal SK, Mishra M, Jain A, Radhakrishnan VB, editors. A closer look at smoothness in domain adversarial training. International Conference on Machine Learning; 2022: PMLR.
33. Jin Y, Wang X, Long M, Wang J, editors. Minimum class confusion for versatile domain adaptation. European Conference on Computer Vision; 2020: Springer.
34. Hoyer L, Dai D, Wang H, Van Gool L, editors. MIC: Masked image consistency for context-enhanced domain adaptation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2023.
35. Ghifary M, Kleijn WB, Zhang M, Balduzzi D, Li W, editors. Deep reconstruction-classification networks for unsupervised domain adaptation. European Conference on Computer Vision; 2016: Springer.
36. Bousmalis K, Silberman N, Dohan D, Erhan D, Krishnan D, editors. Unsupervised pixel-level domain adaptation with generative adversarial networks. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; 2017.
37. Saito K, Ushiku Y, Harada T, Saenko K, editors. Strong-weak distribution alignment for adaptive object detection. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2019.
38. Vayyat M, Kasi J, Bhattacharya A, Ahmed S, Tallamraju R. Cluda: Contrastive learning in unsupervised domain adaptation for semantic segmentation. arXiv preprint arXiv:220814227 2022.
39. Zou Y, Yu Z, Liu X, Kumar B, Wang J, editors. Confidence regularized self-training. Proceedings of the IEEE/CVF International Conference on Computer Vision; 2019.
40. Zhang X, Wu Y, Angelini E, Li A, Guo J, Rasmussen JM, O’Connor TG, Wadhwa PD, Jackowski AP, Li H, editors. Mapseg: Unified unsupervised domain adaptation for heterogeneous medical image segmentation based on 3d masked autoencoding and pseudo-labeling. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2024.
41. Chen L, Chen H, Wei Z, Jin X, Tan X, Jin Y, Chen E, editors. Reusing the task-specific classifier as a discriminator: Discriminator-free adversarial domain adaptation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2022.
42. Ghiasi G, Lin TY, Le QV. Dropblock: A regularization method for convolutional networks. Advances in Neural Information Processing Systems 2018;31.
43. Guo LZ, Li YF, editors. Class-imbalanced semi-supervised learning with adaptive thresholding. International Conference on Machine Learning; 2022: PMLR.
44. Gao B, Qu M, Jiang Y, Li W, Wang M, Pei C, Zheng D, Yang C, Miao Y. Fractional Anisotropy is a More Sensitive Diagnostic Biomarker Than Mean Kurtosis for Patients with Parkinson Disease with Cognitive Dysfunction: A Diffusional Kurtosis Map Tract-Based Spatial Statistics Study. AJNR Am J Neuroradiol 2024;45:1098-105. [Crossref] [PubMed]