Multimodal MRI integrating anti-motion multi-parametric mappings for investigating subcortical nuclei microstructural alterations in Huntington's disease

Abstract

Background

Huntington's disease (HD) is a hereditary neurodegenerative disorder, with pathological changes detectable by MRI before symptom onset. Quantitative MRI (qMRI) provides tissue-specific parameters and holds potential for capturing disease-related biomarkers. However, conventional analysis methods often rely on single-modality imaging or mean features, constraining their ability to capture HD's complex microstructural evolution.

Purpose

To assess the feasibility of multi-modal MRI combined with the MOLED sequence in HD patients and explore its value in early disease detection and staging.

Methods

22 HD patients (14 Pre-HD and 8 M-HD) and 27 healthy controls were enrolled. MOLED-derived T2 and T2* maps, along with structural MRI, were acquired using two 3.0 T scanners to assess inter-scanner consistency. The MOLED sequence incorporates ultrafast acquisition techniques to minimize motion artifacts and improve image quality. Histogram-based features (e.g., variance, skewness, and maximum) and volumes were extracted from eight deep brain regions. Multiple machine learning models were employed for classification analysis.

Results

The MOLED demonstrated good image consistency and reproducibility across scanners. Significant group differences were observed in the volumes of several basal ganglia regions and in variance-based features across multiple modalities. Machine learning models combining clinical and mapping features achieved the highest classification performance (maximum F1-macro = 0.846, Sensitivity-macro = 0.838).

Conclusion

MOLED provides stable and complementary quantitative information for multi-modal MRI. Integrating multimodal multi-feature with machine learning enables a more comprehensive depiction of HD-related microstructural heterogeneity and disease progression.

Plain language summary

Huntington's disease (HD) is a genetic brain disorder that causes gradual damage to nerve cells, often years before any symptoms appear. Detecting these early brain changes is important for better diagnosis and treatment. In this study, we used a special type of MRI scan called MOLED, which can quickly and clearly capture brain images with detailed tissue information. We combined this with standard MRI to examine differences between people with HD (both early-stage and mid-stage) and healthy individuals.

We looked at specific deep brain areas and measured features like shape, size, and signal patterns. Our results showed that the MOLED scan produced reliable images on different MRI machines and could detect meaningful changes in the brain, even in early stages of the disease. When we used machine learning to analyze the data, we were able to accurately distinguish between different groups based on their brain features.

This research shows that combining advanced MRI techniques with computer-based analysis can help detect HD earlier and understand how the disease progresses over time.

Keywords

Huntington's disease multi-modal imaging anti-motion quantitative MRI imaging features machine learning

Introduction

Huntington's disease (HD) is an autosomal dominant neurodegenerative disorder caused by abnormal CAG repeat expansion (≥36) in the HTT gene,^1,2 leading to the expression of mutant huntingtin protein (mHTT) with neurotoxic effects. Accumulation of mHTT progressively impairs neuronal function and ultimately causes cell death.¹ Clinical symptoms typically emerge in the later stages and include progressive motor dysfunction, cognitive decline, and psychiatric disturbances.^3,4 Previous studies have shown that detectable changes in brain anatomy,^5,6 metabolism,⁷ and basal ganglia atrophy^8–10 may occur even before symptom onset.¹⁰ Therefore, identifying early biomarkers is crucial for timely diagnosis, intervention, and understanding disease pathophysiology.³ While genetic testing confirms diagnosis, it lacks sensitivity to disease progression, highlighting the need for reliable neuroimaging biomarkers to monitor disease trajectory and assess therapeutic efficacy.^11,12

Among existing imaging approaches, T1-weighted volumetry is recognized as a reliable biomarker for tracking HD progression.¹³ However, conventional qualitative MRI is limited by inter-scanner variability and subjective interpretation, reducing its sensitivity to subtle tissue changes.¹⁴ In contrast, qMRI provides reproducible and quantifiable tissue parameters,¹⁵ and has shown promise in revealing HD-related pathophysiological mechanisms.¹⁶ For instance, T1ρ imaging has been applied to detect mitochondrial dysfunction^17,18; T2* mapping^14,18 and R2* mapping^19,20 reflect alterations in brain iron or water content; and susceptibility-based imaging techniques such as Magnetic Field Correlation (MFC)²¹ and Quantitative Susceptibility Mapping (QSM)^5,22 further validate endogenous iron accumulation in HD.²³ Moreover, the T1-weighted imaging (T1W)/T2-weighted imaging (T2W) ratio, although influenced by iron deposition and inflammation, serves as a sensitive marker of white matter integrity²⁴ and has been applied to HD research.²⁵

However, conventional qMRI sequences suffer from long acquisition times and high sensitivity to motion,²⁶ posing challenges for HD populations. To address this, MOLED (Multiple Overlapping-Echo Detachment)—a single-shot ultrafast qMRI technique—was developed.²⁷ By enabling multiparametric acquisition in a single excitation,²⁸ MOLED substantially reduces scan time and motion artifacts, enhancing image quality and clinical utility.^26,29 Wu et al.¹⁴ demonstrated MOLED's robustness against intentional head motion and reported significant differences in T2, T2* values, and regions of interest (ROI) volumes between HD patients and healthy controls (HC). Moreover, both ROI volume and T2* values in the frontal white matter were correlated with clinical severity. Despite these advantages, the accuracy of qMRI-derived parameters remains influenced by scanner-specific factors such as hardware configuration and magnetic field strength.³⁰ These systematic variations are especially critical in multi-scanner or multi-center studies, potentially undermining data comparability and reproducibility.³¹

To further uncover the pathological information embedded in MRI data, histogram analysis has gained traction in neuroimaging research.³² Unlike mean-based approaches,¹⁴ histogram-based methods allow the extraction of first-order statistical features such as mean, variance, skewness, and kurtosis from ROIs, providing a more comprehensive assessment of signal heterogeneity and its underlying pathophysiological relevance.^32,33 These features can be further leveraged by machine learning to reduce redundancy, identify key biomarkers,³⁴ improve classification and prediction accuracy.³⁵ Such integrative approaches are increasingly applied in neurodegenerative disease research.^35,36

The methodological framework of this study is illustrated in Figure 1. Quantitative T2²⁷ and T2*³⁷ maps were acquired using the MOLED sequence, alongside structural MRI and the generated T1W/T2W ratio maps. Histogram features and volume features were extracted from multiple subcortical brain regions (e.g., caudate nucleus (Cd), putamen (Put)). To assess the cross-scanner robustness, data were collected on two 3.0 T MRI systems. Statistical analyses were performed to evaluate the diagnostic relevance of individual features. A random forest algorithm was used for feature selection, followed by multi-model classification to assess the utility of mapping, structural, and combined features in detecting HD-related alterations. This framework aims to provide a robust, efficient tool for early diagnosis, disease monitoring, and treatment evaluation in HD. The main contributions of this study include:

（1） Systematic Error Correction: Quantitative consistency across scanners was evaluated, and a U-Net–based correction model was applied to reduce system bias, improving cross-device reproducibility.

（2） Biological Interpretability via T1W/T2W: Integration of T1W/T2W ratio maps, sensitive to iron deposition and neuroinflammation, enhanced the biological relevance of white matter alterations.

（3） Improved Sensitivity through Multimodal, Multi-feature Integration: Combining multiple modalities with histogram and volumetric features enabled a more comprehensive depiction of HD-related microstructural changes across disease stages.

Figure 1.

Image reconstruction and postprocessing pipeline.

Materials and methods

Ethics approval

This study was approved by the First Affiliated Hospital of Zhengzhou University Review Board 2023-KY-1139-002. All participants provided written informed consent prior to scanning.

Patient cohort

This cross-sectional study was approved by the Institutional Review Board of the First Affiliated Hospital of Zhengzhou University.¹⁴ Between April 2023 and May 2024, a total of 36 patients with HD were recruited. Fourteen were excluded due to a CAG repeat length of less than 36 or age ≤ 21 years, yielding a final cohort of 22 HD patients. Based on motor symptomatology, patients were classified into the premanifest HD group (Pre-HD, n = 14) and the manifest HD group (M-HD, n = 8).¹⁴ Additionally, 27 age- and sex-matched HC were enrolled.^38,39 One HC was excluded due to missing MOLED data, resulting in 26 HC participants. The overall participant enrollment process is illustrated in Figure 2, and demographic details are summarized in Table S1.

Figure 2.

Enrollment flowchart.

MRI acquisition

All participants underwent MRI scanning on a 3.0 T system using a 32-channel head coil. All HD patients and 5 HCs were scanned on a Philips Ingenia Elition X scanner (Philips Healthcare, Best, Netherlands), while the remaining 22 HCs were scanned on a Philips Ingenia CX system. The imaging protocol included the following multimodal sequences: T2 MOLED, T2* MOLED, T2W, T1W, and T2-weighted fluid-attenuated inversion recovery (FLAIR). Detailed acquisition parameters are listed in Table S2.

Image preprocessing and registration

A U-Net⁴⁰ network was trained using simulated datasets^37,41 and applied to reconstruct quantitative maps from MOLED acquisitions. To account for inter-scanner variability, separate U-Net models were trained for each MRI system.

T2 and T2* maps were first processed with FSL's brain extraction tool.⁴² Structural images (T1W, T2W, FLAIR) underwent bias-field correction,⁴³ followed by affine registration to the MNI152 standard space. The inverse transformation was applied to project the MNI brain mask into individual space for refined skull stripping. To improve inter-subject consistency, all images were intensity-normalized to the 99th percentile.⁴⁴

As shown in Figure 3B, skull-stripped multimodal images from a HC and a representative HD patient illustrate characteristic changes in HD, including Cd atrophy (dashed arrows), reduced Put signal suggestive of iron deposition (solid arrows), and frontal motion artifacts in structural images (dotted arrows). Notably, the MOLED sequence leverages ultrafast acquisition techniques to mitigate motion artifacts, thereby providing high-quality quantitative T2/T2* maps even in subjects with involuntary movements. Given the higher signal-to-noise ratio and spatial resolution of T2W images, T1W images were registered to T2W space and used to generate T1W/T2W ratio maps.

Figure 3.

Comparison of multimodal MRI between HD and HC. (A) Illustration of eight deep brain region ROIs defined based on the HybraPD atlas. (B) Axial views of T2 mapping, T2* mapping, T1W, T2W, FLAIR, and T1W/T2W from a representative HD subject and a HC. Dashed arrows indicate Cd atrophy; solid arrows highlight signal reduction associated with iron deposition in the Put; dotted arrows denote motion artifacts.

To ensure consistent spatial referencing, both T1W and T2W images of each subject were aligned to their respective mapping spaces using affine and deformable registration methods.⁴⁵ The HybraPD atlas⁴⁶ provides standardized modalities including T1 and R2 images (where R2 = 1/T2), but lacks native T2 images. Accordingly, the T1 and R2 images from the HybraPD atlas were jointly registered to each subject's T1W–T2W space. The resulting transformation matrices were applied to the subcortical segmentation atlases to enable accurate ROI mapping into the native mapping space. For one HC subject lacking T1W data (Sub009HC), FLAIR was used as a substitute for registration. In this case, structural images were registered directly to the T1 modality of the HybraPD atlas, and inverse transformations were used to project ROI labels back to each modality space (Figure 1).

The HybraPD atlas⁴⁶ enables automated segmentation of eight deep brain structures: the Cd, Put, externa/interna globus pallidus (GPe/GPi), thalamus (TH), red nucleus (RN), substantia nigra (SN), and dentate nucleus (DN)—regions frequently implicated in HD pathology.⁵ Although derived from a Parkinson's disease cohort, this atlas was selected as it provides high-quality manual segmentations of the subcortical nuclei most relevant to HD, and a widely accepted HD-specific atlas is currently lacking. To ensure accurate ROI placement despite potential anatomical differences, all resulting masks underwent meticulous visual inspection and manual correction where necessary. The spatial distribution of these subcortical ROIs is illustrated in Figure 3A.

To reduce cerebrospinal fluid (CSF) contamination and partial volume effects in the five MRI modalities and T1W/T2W ratio maps, CSF voxels were excluded using an intensity threshold. ROIs were then subjected to 1-pixel morphological erosion on a slice-by-slice basis. If any ROI retained fewer than four voxels post-erosion, a lower threshold was applied to refine segmentation. All ROI masks were visually inspected and manually edited as needed.

Feature extraction and statistical analysis

Twelve first-order statistical features were extracted from each modality within each ROI, including Mean, Variance, Root Mean Square (RMS), Kurtosis, Skewness, 10th percentile (P10), 25th percentile (P25), median (P50), 75th percentile (P75), 90th percentile (P90), maximum (Max) and minimum (Min). ROI volume was estimated by multiplying the voxel count by the corresponding voxel resolution of the modality.

Cross-Scanner consistency assessment

To evaluate inter-scanner variability, five HCs scanned on the Ingenia Elition X were paired with five age- and sex-matched HCs from the Ingenia CX system (age difference: 0.00 ± 8.25 years, P = 1). In the mapping images, the DN was excluded from analysis due to its proximity to the skull base and limited slice coverage, which led to incomplete representation in some subjects. For each of the five imaging modalities, the mean ROI values were extracted, and the inter-scanner differences were tested for normality. If normality was confirmed, paired t-tests were performed, and agreement between scanners was further evaluated using Bland–Altman plots.

Statistical comparison across groups

For each of the 13 features extracted from eight deep brain ROIs across the HC, Pre-HD, and M-HD groups, normality was assessed using the Shapiro–Wilk test and homogeneity of variance via Levene's test. When both assumptions were met, one-way ANOVA with Tukey's HSD post hoc test was applied. For non-normally distributed variables with equal variances, the Kruskal–Wallis test followed by Dunn's test was used. Statistical significance was defined as P < 0.05.

Classification model development and evaluation

To evaluate the discriminative power of the extracted imaging features in HD diagnosis—specifically those from Mapping (T2, T2*), Clinical (T1W, T2W, FLAIR, T1W/T2W), and Combination (all modalities)—multiple machine learning classifiers were applied. Missing values (e.g., due to absent T1W or dentate nucleus data) were imputed via linear interpolation, and all features were standardized using Z-score normalization to reduce scale variance. Feature selection proceeded in three steps: (1) Correlation filtering (absolute Pearson correlation coefficient > 0.8) to remove redundancy; (2) ANOVA to retain the top 80 features with significant group differences; and (3) LASSO regularization combined with Recursive Feature Elimination (RFE)⁴⁷ using a Random Forest (RF) classifier to identify the final eight features. This dimensionality (approximately 1/6 of the sample size) was selected to minimize overfitting risk. The selected features accounted for approximately one-sixth of the total sample size, thereby preventing overfitting.

Five classifiers—Logistic Regression (LR), RF, Support Vector Machine (SVM), XGBoost, and K-Nearest Neighbors (KNN)—were trained and evaluated using stratified 5-fold cross-validation to ensure class balance and generalizability. Model performance was assessed using macro-averaged F1-score, sensitivity, and specificity, in addition to accuracy and the area under the ROC curve (AUC). The use of macro-averaging ensured an unbiased evaluation across classes of differing sizes.

Results

Instrument consistency assessment

As shown in Figure 4 and Table 1, the two MRI scanners demonstrated good agreement. Paired t-tests revealed no significant differences across any ROI. The lowest P-value was observed in the SN for T2 mapping (−4.04 ± 7.19 ms, P = 0.277), which is greater than the 0.05 significance threshold, indicating no statistically significant inter-scanner difference. Bland–Altman analysis showed mean differences of −2.21 ± 4.91 ms for T2 and 0.05 ± 3.99 ms for T2*, both within acceptable limits. For structural images, the lowest P-values occurred in the TH (T1W: 1.77 ± 1.58%, P = 0.067) and Cd (T2W: 1.42 ± 1.52%, P = 0.106), though neither was significant. Mean inter-scanner differences were 0.60 ± 2.56% for T1W, −0.99 ± 4.50% for T2W, and 0.41 ± 6.21% for FLAIR.

Figure 4.

The consistency analysis of five MRI modalities acquired on two 3.0 T MRI scanners. (A-E) Paired ROI-wise comparisons of T2, T2*, T1W, T2W, and FLAIR values, demonstrating high inter-scanner concordance. (F-J) Bland–Altman plots illustrating the distribution of measurement differences between the two scanners. No statistically significant differences were observed (ns, P > 0.05).

Table 1.

Summarization of the paired statistical analysis of values from five MRI modalities in deep brain regions measured by the two MRI systems.

Modality		Cd	Put	GPe	GPi	TH	RN	SN	DN
T2	error	−1.29 ± 2.73	−1.56 ± 3.67	−1.57 ± 3.52	−0.90 ± 5.85	−3.65 ± 7.24	−2.44 ± 4.75	−4.04 ± 7.19	-
T2	p-value	0.350	0.396	0.374	0.749	0.322	0.315	0.277	-
T2*	error	−1.16 ± 5.10	0.48 ± 2.46	0.82 ± 1.51	1.67 ± 5.16	−0.28 ± 4.08	−2.07 ± 4.28	0.90 ± 5.14	-
T2*	p-value	0.638	0.684	0.294	0.509	0.884	0.340	0.716	-
T1W	error	1.42 ± 1.52	1.16 ± 1.66	2.21 ± 3.56	−0.02 ± 2.84	1.77 ± 1.58	0.27 ± 1.23	−2.47 ± 2.88	0.48 ± 2.74
T1W	p-value	0.106	0.194	0.238	0.988	0.067	0.654	0.127	0.717
T2W	error	−1.46 ± 3.17	−0.77 ± 3.54	−2.59 ± 4.80	−0.16 ± 8.37	−1.21 ± 4.54	−0.02 ± 5.92	−0.55 ± 4.14	−1.20 ± 1.87
T2W	p-value	0.362	0.653	0.294	0.968	0.582	0.995	0.782	0.226
FLAIR	error	−0.10 ± 2.59	−0.03 ± 3.61	−1.18 ± 10.09	3.84 ± 8.05	1.99 ± 3.00	−0.56 ± 3.60	2.29 ± 2.84	−2.96 ± 10.90
FLAIR	p-value	0.936	0.987	0.807	0.347	0.211	0.744	0.146	0.576

Data are presented as mean ± standard deviation; the error is defined as Ingenia CX minus Ingenia Elition X; “–” is used to indicate not applicable. Cd = caudate nucleus; Put = putamen; GPe = globus pallidus externus; GPi = globus pallidus internus; TH = thalamus; RN = red nucleus; SN = substantia nigra; DN = dentate nucleus.

Group differences in T2 and T2* mappings

Only features showing statistically significant intergroup differences are reported. Significance annotations in the figures reflect pairwise post hoc comparisons with Bonferroni correction. Mapping results were presented in Figure 5, Figure S1, Table 2, and Table S3. In the mean values, the Cd significantly decreased in M-HD compared to HC (T2: P < 0.001; T2*: P = 0.024). Variance increased with disease severity, particularly in T2*, the Cd, Put, and TH showed significantly higher variance in M-HD than in HC and Pre-HD (P ≤ 0.037). T2 values showed similar trends in the Put, GPi, and TH (P ≤ 0.047). For lower percentiles (P10, P25, Min), the Cd was significantly lower in M-HD than HC (T2: P ≤ 0.001; T2*: P ≤ 0.003). P50 and P75 of T2 in Cd were lower in M-HD (P ≤ 0.006), while Max in Put and TH increased (P ≤ 0.007). For T2*, the Put showed significantly increased values at higher percentiles (P75, P90, Max) in M-HD (P ≤ 0.027). Volume loss was observed in key regions: T2 volumes of Cd, Put, and GPe decreased with disease progression (P ≤ 0.026), as did T2* volume in Put (P ≤ 0.006).

Figure 5.

Intergroup comparisons of T2* features across ROIs among HD subgroups and HC. Boxplots show feature distributions, with asterisks indicating statistical significance (*p < 0.05, **p < 0.01, *** p < 0.001). Stars denote feature means, which in some cases fall outside the interquartile range, reflecting skewed distributions likely driven by outliers. Notably, skewness in the GPi was significant before but not after multiple comparison correction.

Table 2.

Summarization of the histogram-based feature comparisons of T2* images across HD subgroups and HC for each ROI.

ROI	Feature	HC	Pre_HD	M_HD	HC vs. Pre_HD	HC vs. M_HD	Pre_HD vs. M_HD
Cd	Mean	45.03 ± 3.29	42.71 ± 4.91	40.48 ± 5.25	0.216	0.024	0.449
	Variance	37.70 ± 29.34	58.44 ± 33.83	146.17 ± 61.41	0.081	0.000	0.037
	Skewness	0.50 ± 0.78	0.97 ± 0.82	1.39 ± 0.34	0.197	0.003	0.299
	P10	38.28 ± 4.32	34.17 ± 6.59	27.21 ± 5.83	0.058	0.000	0.013
	P25	41.35 ± 3.88	37.73 ± 6.35	32.20 ± 5.95	0.084	0.000	0.043
	Min	26.75 ± 5.65	23.58 ± 5.07	19.10 ± 4.64	0.183	0.003	0.152
	Max	75.42 ± 14.69	87.20 ± 13.52	116.38 ± 36.70	0.035	0.000	0.200
	Volume	2.96 ± 0.56	2.56 ± 0.51	1.83 ± 1.07	0.164	0.000	0.041
Put	Variance	30.23 ± 19.01	37.86 ± 32.75	108.49 ± 58.78	1.000	0.000	0.002
	P75	45.82 ± 2.78	46.35 ± 3.73	50.48 ± 5.01	0.892	0.005	0.027
	P90	48.87 ± 2.56	49.48 ± 3.08	55.63 ± 4.83	0.830	0.000	0.000
	Min	26.46 ± 5.30	27.62 ± 6.64	18.41 ± 8.08	1.000	0.027	0.014
	Max	60.10 ± 7.99	62.68 ± 10.97	84.00 ± 17.12	1.000	0.000	0.004
	Volume	5.30 ± 0.98	4.07 ± 1.01	2.64 ± 0.93	0.001	0.000	0.006
GPe	Variance	24.10 ± 8.82	43.07 ± 20.22	75.64 ± 44.26	0.009	0.002	1.000
	Skewness	0.33 ± 0.46	0.43 ± 0.43	0.94 ± 0.58	0.800	0.006	0.046
	P10	23.50 ± 4.68	22.06 ± 4.02	17.83 ± 3.02	0.569	0.005	0.076
	P25	26.04 ± 4.74	25.02 ± 4.62	20.51 ± 4.09	0.780	0.013	0.081
	Min	19.30 ± 4.39	18.09 ± 3.10	14.80 ± 2.02	0.595	0.013	0.132
	Volume	1.16 ± 0.24	0.88 ± 0.23	0.64 ± 0.22	0.011	0.000	0.283
GPi	Variance	11.47 ± 6.07	23.14 ± 19.48	49.84 ± 28.94	0.283	0.003	0.227
	Skewness	0.59 ± 0.60	0.96 ± 0.73	1.25 ± 0.91	0.259	0.059	0.614
	P10	19.98 ± 4.75	18.62 ± 3.47	15.33 ± 1.78	1.000	0.018	0.130
	Min	17.70 ± 4.48	16.67 ± 3.34	13.73 ± 1.14	1.000	0.023	0.106
	Max	33.67 ± 5.85	38.69 ± 10.73	44.41 ± 11.48	0.350	0.012	0.442
	Volume	0.44 ± 0.09	0.37 ± 0.07	0.28 ± 0.05	0.021	0.000	0.050
TH	Variance	20.41 ± 6.58	22.21 ± 8.02	42.73 ± 18.11	1.000	0.001	0.009
TH	P10	43.78 ± 2.56	43.12 ± 3.28	40.26 ± 4.10	0.788	0.017	0.098
RN	Variance	19.39 ± 12.75	28.84 ± 27.11	68.73 ± 60.68	1.000	0.006	0.087
	P10	27.67 ± 4.91	29.62 ± 6.00	23.16 ± 3.90	0.483	0.083	0.017
	P25	29.84 ± 4.77	31.77 ± 6.43	25.50 ± 4.06	0.500	0.107	0.024
	Min	25.14 ± 4.37	27.23 ± 6.10	21.38 ± 4.20	0.402	0.149	0.026
	Max	42.01 ± 4.92	46.33 ± 6.90	48.86 ± 10.91	0.136	0.039	0.675
SN	Variance	29.24 ± 13.81	41.06 ± 20.53	61.27 ± 33.85	0.137	0.023	1.000
	P10	23.45 ± 5.03	25.18 ± 5.35	19.17 ± 3.10	1.000	0.043	0.011
	P25	25.94 ± 4.78	28.17 ± 5.91	22.19 ± 3.54	0.371	0.157	0.025
	Min	20.82 ± 4.65	21.82 ± 5.17	16.75 ± 3.97	1.000	0.074	0.046
	Max	42.63 ± 5.75	48.85 ± 7.97	46.58 ± 8.34	0.040	0.492	1.000
DN	Mean	36.60 ± 4.78	37.34 ± 6.84	30.15 ± 9.11	0.938	0.055	0.048
	RMS	37.02 ± 4.66	37.71 ± 6.75	30.45 ± 9.03	0.944	0.046	0.041
	P75	40.01 ± 4.87	40.83 ± 6.95	32.40 ± 9.06	0.924	0.021	0.018
	P90	43.12 ± 5.03	44.16 ± 6.62	36.02 ± 9.25	0.885	0.034	0.024
	Max	49.64 ± 7.47	49.50 ± 6.71	41.15 ± 7.34	0.998	0.025	0.047

Comparisons with multiple comparison corrected p-values < 0.05 are highlighted in bold font. Values are represented by mean ± standard deviation.

Group differences in clinical images

Clinical image analysis results are presented in Figure S2-S4 and Table S4-S6. In T2W images, the mean intensities of the Put, GPe, GPi, and TH were significantly lower in the M-HD (P ≤ 0.006). Variance increased in M-HD across T1W (Put, GPe, GPi, TH; P ≤ 0.030), T2W (Cd, Put; P ≤ 0.033), and FLAIR (Cd, Put, TH; P ≤ 0.025). Skewness and kurtosis were elevated in the M-HD group for the Put in T2W (P ≤ 0.004). Lower percentiles (P10, P25, Min) were reduced in M-HD, particularly in the Put, GPe, and TH in T2W (P ≤ 0.038), and the Cd in FLAIR (P ≤ 0.022). P50 values in T2W were significantly lower in the Put, GPe, and TH in M-HD (P ≤ 0.009). Higher percentiles (P75, P90, Max) were significantly increased in the GPi in T2W (P ≤ 0.022). Volume analysis revealed a pattern of structural atrophy consistent with that observed in mapping data. The Cd and TH volumes were significantly reduced in the M-HD group compared to HC and Pre-HD in both T1W (P ≤ 0.030) and FLAIR (P ≤ 0.041).

Group differences in T1W/T2W

In the analysis of T1W/T2W ratio maps (Figure 6, Table S7), both mean and RMS values were significantly higher in the M-HD group compared to HC and Pre-HD in the Put, GPe, and GPi (P ≤ 0.047). Variance was significantly increased in the Put, GPe, TH, red nucleus (RN), and substantia nigra (SN) in M-HD (P ≤ 0.023). For lower to mid percentiles (P10, P25, P50), the GPi showed significantly higher values in M-HD compared to HC (P ≤ 0.013). Higher percentiles (P75, P90) were significantly elevated in the Put, GPe, GPi, and SN in M-HD compared to both HC and Pre-HD (P ≤ 0.041).

Figure 6.

Intergroup comparisons of T1W/T2W features across ROIs among HD subgroups and HC. Boxplots show feature distributions, with asterisks indicating statistical significance (*p < 0.05, **p < 0.01, *** p < 0.001).

Machine learning classifiers

As shown in Figure 7 and Table 3, the KNN model achieved the best performance using mapping features (F1-macro = 0.738, sensitivity-macro = 0.712, AUC = 0.8610). For clinical features, performance was comparable across classifiers, with LR performing best (F1-macro = 0.756, sensitivi ty-macro = 0.743, AUC = 0.9356), indicating that mapping and clinical features provide similar classification capacity for HD staging. Combining both modalities further improved performance, with the RF model achieving the highest F1-macro (0.846), sensitivity-macro (0.838), accuracy (0.857), and AUC (0.9071), highlighting the benefit of multimodal integration.

Figure 7.

Classification performance of machine learning models based on features from different imaging modalities. (A–C) Comparisons of ROC curves and AUC values using features from mapping, Clinical, and their combination, respectively. (D–F) The rankings of the most contributive features under each condition.

Table 3.

Summarization of the corresponding classification metrics.

Parameters		LR	RF	SVM	XGBoost	KNN
Mapping	F1-Macro (%)	0.701	0.688	0.730	0.627	0.738
	Sensitivity-Macro (%)	0.683	0.670	0.702	0.593	0.712
	Specificity-Macro (%)	0.849	0.840	0.841	0.799	0.837
	Accuracy (%)	0.735	0.714	0.755	0.653	0.735
Clinical	F1-Macro (%)	0.756	0.750	0.691	0.681	0.687
	Sensitivity-Macro (%)	0.743	0.766	0.678	0.670	0.665
	Specificity-Macro (%)	0.893	0.885	0.874	0.843	0.857
	Accuracy (%)	0.813	0.792	0.771	0.729	0.750
Combination	F1-Macro (%)	0.774	0.846	0.811	0.825	0.790
	Sensitivity-Macro (%)	0.755	0.838	0.797	0.821	0.773
	Specificity-Macro (%)	0.893	0.912	0.903	0.919	0.888
	Accuracy (%)	0.816	0.857	0.837	0.857	0.816

The best-performing results are displayed in bold.

Feature importance analysis revealed that, in the mapping model, the most discriminative features were T2*_Put_Volume, T2*_Cd_Max, and T2*_Cd_Skewness. In the Clinical model, T1W_DN_Max, FLAIR_GPe_Variance, and T1W_TH_Volume contributed most. In the Combined model, FLAIR_GPe_Variance, T2_TH_Skewness consistently ranked among the top features, underscoring their pivotal role in multimodal integration.

Discussion

This study employed MOLED-derived T2 and T2* quantitative maps, alongside conventional structural MRI, to investigate iron deposition and microstructural alterations in subcortical regions across HD stages using histogram and volume features. Compared to structural MRI, which is more susceptible to motion artifacts, MOLED T2 and T2* maps maintain image quality even in the presence of involuntary movements. Consistency analyses demonstrated the comparability and reproducibility of MOLED maps. Our results suggest: (1) In the Pre-HD stage, multiple underlying pathological mechanisms may interact, potentially “offsetting” or masking imaging signal changes; (2) With disease progression, subcortical signal distributions exhibit increased variance and polarization, reflecting structural disorganization; (3) MOLED provides complementary information to structural MRI, and—by leveraging ultrafast acquisition for inherent motion resistance—enables more comprehensive characterization of HD-related structural and microstructural alterations. Notably, our prior work¹⁴ has established significant correlations between MOLED-derived features and clinical severity metrics (UHDRS, CAG), supporting their biological relevance.

Modern MRI scanners utilize automatic parameter optimization,⁴⁸ where even slight variations in TE can affect quantitative accuracy—particularly in multi-echo sequences like MOLED. Our findings demonstrate that modality-specific simulation combined with U-Net–based correction can effectively eliminate scanner-induced bias. For structural images, post-processing techniques such as bias field correction and intensity normalization further reduce inter-scanner variability. Owing to its ultrafast acquisition and intrinsic motion resistance, MOLED not only reduces dependency on subject compliance but also enhances its feasibility and reliability for multicenter and longitudinal HD studies.⁶

In the Pre-HD stage, multiple imaging-derived features demonstrate early alterations, reflecting microscopic neuropathological changes preceding overt clinical symptoms.⁴⁹ Volume reductions in the Cd, Put, and GP suggest early neuronal loss,⁷ consistent with previous findings.^5,14 Increased T2* variance in the GPe reflects enhanced tissue heterogeneity, potentially indicating early iron deposition.⁵⁰ A reduction in Skewness in the TH from the T1W/T2W suggests a shift from right-skewed intensity distribution, possibly due to increased oligodendrocyte numbers and myelin content.^6,25 In FLAIR images, elevated P90 in the Cd and Max in the Put may reflect neuroinflammation, supported by astrocytic and oligodendrocytic activation.^1,51 Furthermore, selective degeneration of basal ganglia white matter tracts may contribute to these FLAIR signal changes.⁶ Metabolic disturbances have also been reported in Pre-HD, including reduced glucose uptake, lactate accumulation, and subsequent acidosis, indicating early mitochondrial dysfunction.^16,18 Cerebral blood flow may increase compensatorily to sustain energy supply in the early stages.⁵²

As the disease progresses to the M-HD stage, compensatory mechanisms diminish, and widespread, consistent pathological changes emerge. Imaging reveals progressive and global atrophy alongside more prominent neuronal loss.^5,14 Variance and Skewness increase markedly, particularly in the Put and TH on T2* maps, with a broader signal distribution reflected by declining low percentiles and rising high percentiles. Iron accumulation intensifies, further lowering MRI signal intensity,^21,25 while myelin loss may lead to local signal elevation.⁶ Excessive iron also disrupts mitochondrial function and alters neurotransmitter systems, such as reduced dopamine receptor density, compounding both structural and functional damage.^19,49 In parallel, Cerebral blood flow declines, indicating impaired neurovascular regulation and weakened energetic responses to metabolic demands.⁵² Altogether, imaging alterations across HD stages likely reflect a convergence of microstructural degeneration and physiological dysfunction. While the T1W/T2W ratio may reflect a combination of iron, myelin, and inflammatory changes, its co-occurrence with T2* reduction and atrophy supports a predominant interpretation of iron- and myelin-related microstructural disruption in HD.

In this study, volumetric analysis was performed using multimodal MRI data, rather than relying solely on conventional T1W images. This approach was motivated by the use of an accelerated MRI protocol with a limited number of slices (21) and relatively large slice thickness, which may introduce measurement bias—especially for small deep nuclei such as the RN and SN. Moreover, inter-modality differences in slice thickness and inter-slice spacing, coupled with the small size of the ROIs, increased sensitivity to resolution and motion artifacts.⁵³ These factors likely contributed to the minor inconsistencies observed across modalities, with motion artifacts particularly affecting some T1W images and compromising volumetric accuracy. By integrating information from multiple modalities, the Combination strategy helps mitigate biases inherent in individual sequences due to suboptimal acquisition parameters or image quality.

From a classification perspective, features derived from Mapping—including Variance, Skewness, extreme values, and Volume—effectively captured pathological alterations related to tissue heterogeneity and signal variability in subcortical regions. Despite their lower spatial resolution, the quantitative nature of these maps renders them less susceptible to confounding factors, achieving classification performance comparable to that of Clinical modalities. Importantly, Combination features significantly improved model performance, highlighting the complementary strengths of structural and quantitative imaging in capturing the heterogeneity of HD. Feature importance analysis further confirmed the consistent contribution of volume and variance across multiple models, suggesting their robust potential as imaging biomarkers for HD staging. Among all regions, the Cd and Put were identified as key contributors. These results underscore the utility of MOLED as a valuable supplement to structural MRI, offering improved sensitivity for early HD detection and staging.

Voxel-based morphometry (VBM) offers whole-brain coverage and fully automated analysis, thereby avoiding the subjective bias introduced by manual ROI selection.^54,55 In this study, we implemented a VBM pipeline involving registration to the PD_T1 template, resampling to 1 × 1 × 1 mm isotropic resolution, and 6 mm FWHM spatial smoothing. Nonparametric permutation testing was conducted using the Randomise tool with Threshold-Free Cluster Enhancement (TFCE) correction (p < 0.05). As shown in Figure S5, among the quantitative mappings, only T2 exhibited mild significance in the frontal lobe and Cd when comparing HC and Pre-HD. In contrast, structural images—particularly T1W and FLAIR—revealed widespread differences in M-HD versus other groups. This discrepancy may reflect several factors: (1) the original dataset contained only 21 slices, and interpolation plus deformation-based registration may have introduced spatial inaccuracies⁵⁶; (2) motion artifacts in structural images could have been amplified during spatial normalization; and (3) the T1-based template may have impaired registration accuracy for quantitative maps due to contrast mismatch, thereby reducing statistical sensitivity.⁵⁷ Motion artifacts and multimodal contrast differences reduce the sensitivity and specificity of VBM analyses, highlighting its challenges and supporting ROI-based methods for detailed subcortical assessment in HD.

Limitations and future directions

This study has several limitations. First, the study is based on a relatively small, single-center sample. Additional healthy participants are needed to systematically assess inter-scanner repeatability and model generalizability. Moreover, future multi-center studies with larger and more heterogeneous HD cohorts are crucial for validating the current findings and enhancing the model's robustness and clinical applicability. Second, the absence of direct pathological validation limits mechanistic interpretations. Current inferences regarding iron deposition and myelin changes rely primarily on MRI signal sensitivity, potentially confounded by non-specific factors such as inflammation or cerebral perfusion. Future studies should incorporate complementary biomarkers—such as cerebrospinal fluid analysis, PET imaging, or histopathological evidence—to enhance pathological specificity. Third, while the early-feature-concatenation approach adopted here enhances robustness against overfitting given the sample size, it represents an initial step toward true multimodal fusion. Future studies with larger cohorts will enable the application of more sophisticated frameworks—such as attention-based integration, cross-modal correlation analysis, or graph neural networks—to better capture nonlinear inter-modal interactions and uncover latent pathological biomarkers.

Conclusion

This study combined MOLED-based T2/T2* mapping with multimodal MRI analysis to characterize microstructural and iron-related changes across HD stages. MOLED showed strong inter-scanner consistency and motion resistance, improving image reliability. Key features, especially variance, demonstrated high discriminative power for disease classification. By integrating Mapping, Clinical features, and machine learning, the proposed framework improved sensitivity to subtle disease-related alterations. These findings underscore the potential of MOLED as a robust complement to conventional MRI for early detection and staging of Huntington's disease.

Supplemental Material

sj-docx-1-hun-10.1177_18796397251411608 - Supplemental material for Multimodal MRI integrating anti-motion multi-parametric mappings for investigating subcortical nuclei microstructural alterations in Huntington's disease

Supplemental material, sj-docx-1-hun-10.1177_18796397251411608 for Multimodal MRI integrating anti-motion multi-parametric mappings for investigating subcortical nuclei microstructural alterations in Huntington's disease by Mengying Zhu, Ming Ye, Zejun Wu, Jianzhong Lin, Fei Wu, Xiao Wang, Haiyang Luo, Yong Zhang, Jianfeng Bao, Shuhui Cai and Congbo Cai in Journal of Huntington's Disease

Footnotes

Acknowledgements

The authors have no acknowledgements to report.

ORCID iDs

Mengying Zhu

Shuhui Cai

Congbo Cai

Ethical considerations

This study was approved by the First Affiliated Hospital of Zhengzhou University Review Board 2023-KY-1139-002.

Consent to participate

All participants provided written informed consent prior to scanning.

Consent to publication

All authors consent to the publication of this article.

Author contributions

Mengying Zhu performed the experiments and drafted the original manuscript.

Ming Ye and Zejun Wu provided technical guidance on experimental procedures and MOLED processing.

Jianzhong Lin supervised the image registration methodology.

Fei Wu, Xiao Wang, Haiyang Luo, and Yong Zhang assisted with data collection.

Jianfeng Bao was responsible for data provision and project coordination.

Congbo Cai and Shuhui Cai supervised the project, acquired funding, and revised the manuscript.

All authors reviewed the results and approved the final version of the manuscript.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Natural Science Foundation of China, the Key Technologies R&D Program of Henan Province, (grant number 12375291, 62331021, 82071913, 242102311243).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability

Data collected as part of this study will be made available upon reasonable request.

Supplemental material

Supplemental material for this article is available online.

References

Tabrizi

Estevez-Fraga

Van Roon-Mom

WMC

, et al. Potential disease-modifying therapies for Huntington’s disease: lessons learned and future opportunities. Lancet Neurol 2022; 21: 645–658.

Lila

Hunt

Child

, et al. Asymmetric brain atrophy in Huntington’s disease: a postmortem MRI study. J Huntingt Dis 2025; 14: 132–139.

Tabrizi

Schobel

Gantman

, et al. A biological classification of Huntington’s disease: the integrated staging system. Lancet Neurol 2022; 21: 632–644.

Caillaud

Laisney

Bejanin

, et al. Social cognition profile in early Huntington disease: insight from neuropsychological assessment and structural neuroimaging. J Huntingt Dis 2024; 13: 467–477.

Yao

Morrison

Jakary

, et al. Altered iron and microstructure in Huntington’s disease subcortical nuclei: insight from 7 T MRI. J Magn Reson Imaging 2024; 60: 1484–1499.

Estevez-Fraga

Scahill

Rees

, et al. Diffusion imaging in Huntington’s disease: comprehensive review. J Neurol Neurosurg Psychiatry 2021; 92: 62–69.

Bao

Xie

, et al. Application value of Serum neurofilament light protein for disease staging in Huntington’s disease. Mov Disord 2023; 38: 1307–1315.

Postert

Lack

Kuhn

, et al. Basal ganglia alterations and brain atrophy in Huntington’s disease depicted by transcranial real time sonography. J Neurol Neurosurg Psychiatry 1999; 67: 457–462.

Leitão

Guerreiro

Nunes

, et al. Neuromelanin magnetic resonance imaging of the substantia Nigra in Huntington’s disease. J Huntingt Dis 2020; 9: 143–148.

10.

Van Wamelen

Martin

Makos

, et al. Study protocol for the iMarkHD study in individuals with Huntington’s disease. J Huntingt Dis 2024; 13: 479–489.

11.

Yao

Morrison

Jakary

, et al. Altered iron and microstructure in Huntington’s disease subcortical nuclei: insight from 7 T MRI. J Magn Reson Imaging 2024; 60: 1484–1499.

12.

Aqel

Ahmad

Saleh

, et al. Advances in Huntington’s disease biomarkers: a 10-year bibliometric analysis and a comprehensive review. Biology (Basel) 2025; 14: 129.

13.

Pérot

J-B

Brouillet

Flament

. The contribution of preclinical magnetic resonance imaging and spectroscopy to Huntington’s disease. Front Aging Neurosci 2024; 16: 1306312.

14.

Luo

Wang

, et al. Application of anti-motion ultra-fast quantitative MRI in neurological disorder imaging: insights from Huntington’s disease. J Magn Reson Imaging 2025; 61: 2455–2468.

15.

Weiskopf

Edwards

Helms

, et al. Quantitative magnetic resonance imaging of brain anatomy and in vivo histology. Nat Rev Phys 2021; 3: 570–588.

16.

Fazio

Paucar

Svenningsson

, et al. Novel imaging biomarkers for Huntington’s disease and other hereditary Choreas. Curr Neurol Neurosci Rep 2018; 18: 85.

17.

Tereshchenko

Schultz

Kunnath

, et al. Subcortical T1-rho MRI abnormalities in juvenile-onset Huntington’s disease. Brain Sci 2020; 10: 533.

18.

Wassef

Wemmie

Johnson

, et al. T1ρ imaging in premanifest Huntington disease reveals changes associated with disease progression. Mov Disord 2015; 30: 1107–1114.

19.

Sánchez-Castañeda

Cherubini

Elifani

, et al. Seeking huntington disease biomarkers by multimodal, cross-sectional basal ganglia imaging. Hum Brain Mapp 2013; 34: 1625–1635.

20.

Sánchez-Castañeda

Squitieri

Di Paola

, et al. The role of iron in gray matter degeneration in Huntington’s disease: a magnetic resonance imaging study: the role of iron in gray matter loss in HD. Hum Brain Mapp 2015; 36: 50–66.

21.

Dumas

Versluis

Van Den Bogaard

SJA

, et al. Elevated brain iron is independent from atrophy in Huntington’s disease. NeuroImage 2012; 61: 558–564.

22.

Yao

Morrison

Jakary

, et al. Comparison of quantitative susceptibility mapping methods for iron-sensitive susceptibility imaging at 7T: an evaluation in healthy subjects and patients with Huntington’s disease. NeuroImage 2023; 265: 119788.

23.

Chen

Hardy

Kucharczyk

, et al. MR Of human postmortem brain tissue: correlative study between T2 and assays of iron and ferritin in Parkinson and Huntington disease. AJNR Am J Neuroradiol 1993; 14: 275–281.

24.

Tullo

Patel

Devenyi

, et al. MR-Based age-related effects on the striatum, globus pallidus, and thalamus in healthy individuals across the adult lifespan. Hum Brain Mapp 2019; 40: 5269–5288.

25.

Rowley

Tabrizi

Scahill

, et al. Altered intracortical T1-weighted/T2-weighted ratio signal in Huntington’s disease. Front Neurosci 2018; 12: 805.

26.

Dai

Yang

Zhou

, et al. Single-shot T₂ mapping via multi-slice information sharing based on switching modulation patterns multiple overlapping-echo detachment imaging. Med Phys 2025; 52: 4464–4479.

27.

Zhang

Chen

, et al. Robust single-shot T₂ mapping via multiple overlapping-Echo acquisition and deep neural network. IEEE Trans Med Imaging 2019; 38: 1801–1811.

28.

Yang

, et al. Single-shot multi-parametric mapping based on multiple overlapping-echo detachment (MOLED) imaging. NeuroImage 2022; 263: 119645.

29.

Yang

Lin

Wang

, et al. MOdel-Based SyntheTic data-driven learning (MOST-DL): application in single-shot T₂ mapping with severe head motion using overlapping-Echo acquisition. IEEE Trans Med Imaging 2022; 41: 3167–3181.

30.

Nowak

Henningsson

Davis

, et al. Repeatability, reproducibility, and observer variability of cortical T1 mapping for renal tissue characterization. J Magn Reson Imaging 2025; 61: 1914–1922.

31.

Fujita

Gagoski

Nielsen

, et al. Vendor-agnostic 3D multiparametric relaxometry improves cross-platform reproducibility. Magn Reson Med 2025; 94: 937–948.

32.

Zhang

Wang

, et al. Preoperative subtyping of WHO grade 1 meningiomas using a single-shot ultrafast MR T2 mapping. J Magn Reson Imaging 2024; 60: 964–976.

33.

Wang

Zhang

, et al. Positive progesterone receptor expression in meningioma may increase the transverse relaxation: first prospective clinical trial using single-shot ultrafast T2 mapping. Acad Radiol 2024; 31: 187–198.

34.

Zhang

, et al. Multi-Feature extraction and selection method to diagnose burn depth from burn images. Electronics (Basel) 2024; 13: 3665.

35.

Biswas

Gini

. Multi-class classification of Alzheimer’s disease detection from 3D MRI image using ML techniques and its performance analysis. Multimed Tools Appl 2023; 83: 33527–33554.

36.

Trejo-Castro

Caballero-Luna

Garnica-López

, et al. Signal and texture features from T2 maps for the prediction of mild cognitive impairment to Alzheimer’s disease progression. Healthcare 2021; 9: 941.

37.

Yang

Zhou

, et al. Rapid high-fidelity T2* mapping using single-shot overlapping-echo acquisition and deep learning reconstruction. Magn Reson Med 2023; 89: 2157–2170.

38.

Roy

Slimani

De Meester

, et al. Age and sex corrected normal reference values of T1, T2 T2* and ECV in healthy subjects at 3 T CMR. J Cardiovasc Magn Reson 2016; 19: 72.

39.

Kumar

Delshad

Woo

, et al. Age-related regional brain T2-relaxation changes in healthy adults. J Magn Reson Imaging 2012; 35: 300–308.

40.

Ronneberger

Fischer

Brox

. U-Net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, 2015, Vol. 9351, pp.234–241.

41.

Ouyang

Yang

Wang

, et al. Single-shot T₂ mapping via multi-echo-train multiple overlapping-echo detachment planar imaging and multitask deep learning. Med Phys 2022; 49: 7095–7107.

42.

Jenkinson

Beckmann

Behrens

TEJ

, et al. FSL. NeuroImage 2012; 62: 782–790.

43.

Tustison

Avants

Cook

, et al. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging 2010; 29: 1310–1320.

44.

Kascenas

Sanchez

Schrempf

, et al. The role of noise in denoising models for anomaly detection in medical images. Med Image Anal 2023; 90: 102963.

45.

Avants

Tustison

Song

, et al. A reproducible evaluation of ANTs similarity metric performance in brain image registration. NeuroImage 2011; 54: 2033–2044.

46.

Guan

, et al. HybraPD atlas: towards precise subcortical nuclei segmentation using multimodality medical images in patients with Parkinson disease. Hum Brain Mapp 2021; 42: 4399–4421.

47.

Xia

Yang

. A model-free feature selection technique of feature screening and random forest-based recursive feature elimination. Int J Intell Syst 2023; 2023: 2400194.

48.

Mirzaalian

Ning

Savadjiev

, et al. Inter-site and inter-scanner diffusion MRI data harmonization. NeuroImage 2016; 135: 311–323.

49.

Tabrizi

Schobel

Gantman

, et al. A biological classification of Huntington’s disease: the integrated staging system. Lancet Neurol 2022; 21: 632–644.

50.

Chen

Hua

Ross

, et al. Altered brain iron content and deposition rate in Huntington’s disease as indicated by quantitative susceptibility MRI. J Neurosci Res 2019; 97: 467–479.

51.

Van De Zande

Bulk

Najac

, et al. Study protocol of IMAGINE-HD: imaging iron accumulation and neuroinflammation with 7T-MRI + CSF in Huntington’s disease. NeuroImage Clin 2023; 39: 103450.

52.

Liu

Zhang

, et al. Huntingtin silencing delays onset and slows progression of Huntington’s disease: a biomarker study. Brain 2021; 144: 3101–3113.

53.

Sun

Wang

, et al. A foundation model for enhancing magnetic resonance images and downstream segmentation, registration and diagnostic tasks. Nat Biomed Eng 2024; 9: 521–538.

54.

Wang

, et al. Gray matter alterations in Huntington’s disease: a meta-analysis of VBM neuroimaging studies. J Neurosci Res 2024; 102: e25366.

55.

Tan

Shishegar

Poudel

, et al. Cortical morphometry and neural dysfunction in Huntington’s disease: a review. Eur J Neurol 2021; 28: 1406–1419.

56.

Q-H

. Multiscale fusion for spatially decoupled multimodal MRI super-resolution reconstruction. IEEE Trans Instrum Meas 2025; 74: 1–12.

57.

Zhao

Huang

Jiang

, et al. Accurate registration of cross-modality geometry via consistent clustering. IEEE Trans Vis Comput Graph 2024; 30: 4055–4067.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

11.92 MB

0.00 MB