Voxel‐Wise Whole‐Brain Analysis in Quantitative Susceptibility Mapping: A Narrative Review

ABSTRACT

Quantitative susceptibility mapping (QSM) is a post‐processing magnetic resonance imaging technique that quantifies the magnetic susceptibility of biological tissue and provides insights into factors such as iron deposition, hemorrhage, calcification, myelin content, and oxygen extraction fraction. A variety of analytical approaches have been developed to interpret QSM data from different perspectives, including region‐of‐interest‐based, depth‐wise, surface‐based, network‐based, and voxel‐wise methods. Among these, voxel‐wise analysis has gained increasing prominence due to its ability to perform a detailed examination of the entire brain without anatomically predefined regions. This approach is especially valuable for investigating neurological pathologies and aging‐related changes, including neurodegenerative and neuropsychiatric disorders. This article aims to comprehensively summarize voxel‐wise analysis in QSM by outlining key methodological considerations and clinical applications. Moreover, it offers practical data processing recommendations to advance the reproducibility and transparency of voxel‐wise QSM research.

Key Points

• This review presents an overview of voxel‐wise whole‐brain quantitative susceptibility mapping (QSM) with a particular focus on its end‐to‐end analytical pipeline, highlighting post‐processing workflow, methodological considerations, and clinical applications that underlie this technique.

• Voxel‐wise QSM serves as a valuable tool for investigating aging and neurological disorders, as it enables the detection of subtle magnetic susceptibility changes and reveals disease associations without requiring predefined regions of interest.

• Voxel‐wise QSM results depend critically on multiple technical factors, including reconstruction algorithms and post‐processing steps, all of which influence statistical outcomes.

Voxel‐wise QSM enables the detection of subtle susceptibility changes and the exploration of disease associations without relying on predefined regions, making it a valuable tool for studying aging and neurological disorders. This review summarizes the complete workflow of voxel‐wise QSM, highlighting post‐processing steps, key methodological considerations, and clinical applications.

1. Introduction

Magnetic susceptibility reflects the degree of magnetization a substance gains when exposed to a magnetic field, with the magnitude of magnetization proportional to the intensity of the magnetic field (Schenck 1996; Mikl et al. 2008). In the human brain, magnetic susceptibility arises from both paramagnetic substances (such as iron in ferritin and deoxyhemoglobin) and diamagnetic substances (such as myelin and calcification) (Duyn 2013; Wang and Liu 2015). The spatial distribution and concentration of these substances are critical for neurological function and are involved in various pathophysiological processes, including aging (Hallgren and Sourander 1958) and neurodegeneration (Rouault 2013).

Quantitative susceptibility mapping (QSM) is a non‐invasive magnetic resonance imaging (MRI) technique that sensitively visualizes and quantifies the spatial distribution of magnetic susceptibility (Haacke et al. 2015; Sood et al. 2017; Langkammer et al. 2012; De Rochefort et al. 2008, 2010). This sensitivity allows QSM to provide unique insights into both pathological and physiological processes, offering valuable information for disease diagnosis, monitoring disease progression and treatment response. The clinical and research applications of QSM have expanded rapidly (Verma et al. 2022; Vinayagamani et al. 2021; Fushimi et al. 2024; Wang et al. 2017), including neurodegenerative diseases such as Parkinson's disease (PD) and related Parkinsonian syndromes (Li et al. 2019; Zhao et al. 2023; Mohammadi and Ghaderi 2024), mild cognitive impairment (MCI) and Alzheimer's disease (AD) (van Bergen et al. 2016; Paul et al. 2024), amyotrophic lateral sclerosis (ALS) (Schweitzer et al. 2015; Bhattarai et al. 2022), Wilson's disease (WD) (Fritzsch et al. 2014; Li, Wu, et al. 2020), and Huntington's disease (HD) (Domínguez et al. 2016; Yao et al. 2023), as well as neurological disorders such as pediatric Tourette syndrome (TS) (Lin et al. 2025). Additionally, QSM has been applied to traumatic brain injury (TBI) (Liu et al. 2016, 2019), metabolic brain disorders such as hepatic encephalopathy (HE) (Lan et al. 2024), cerebrovascular diseases like stroke (Probst et al. 2021), and infectious diseases such as COVID‐19 (Rua et al. 2024). Furthermore, QSM has been explored in psychiatric and oncological conditions, including major depressive disorder (Liang et al. 2024) and brain tumors (Zeng et al. 2023), and has proven valuable in neurosurgical procedures like deep brain stimulation (DBS) placement (Dimov et al. 2018; Wei et al. 2019).

QSM decodes tissue‐associated magnetic susceptibility, and a series of post‐processing steps are required to reconstruct QSM images from phase and magnitude images (Wang and Liu 2015; Bilgic et al. 2024). Following QSM reconstruction, a crucial step involves comprehensive data analysis to extract multidimensional pathophysiological insights. Various quantitative analysis methods have been developed (Figure 1), including region‐of‐interest (ROI)‐based, depth‐wise, surface‐based, network‐based, and voxel‐wise approaches. ROI‐based approach averages magnetic susceptibility values within predefined regions at the participant level, providing a simple and efficient means of analysis (Figure 1a). Many studies have targeted deep gray matter nuclei (Zhao et al. 2023; Wu et al. 2023; Raab et al. 2022; Zhang et al. 2021; Ward et al. 2019; Darki et al. 2016), which are well‐documented sites of iron deposition (Hallgren and Sourander 1958). However, it is limited by its inability to account for the heterogeneity within regions, and is susceptible to volume effects, which may lead to the loss of local detail. Depth‐wise method (Figure 1b) segments the cortex by its depth, providing insights into cortical microstructure and function (Northall et al. 2023, 2024; Bulk et al. 2020; Merenstein et al. 2024), though correspondence between the defined ‘layers’ and the true biological layers remains uncertain. Similarly, surface‐based method (Figure 1c) maps magnetic susceptibility values onto a surface template (Li, Jacob, et al. 2023), enhancing its applicability in cortical studies. However, cortical erosion inherent to QSM reconstruction may lead to partial loss of cortical information. Network‐based approach (Figure 1d) detects connectivity patterns across brain regions using independent component analysis (ICA) (Reeves et al. 2022) or Kullback–Leibler divergence similarity estimation (KLSE) (Chen et al. 2024). However, it is sensitive to ICA parameters and may also be limited by cortical erosion.

FIGURE 1.

Overview of various methods for QSM analysis. (a) Region‐of‐interest (ROI)‐based analysis examines magnetic susceptibility values within a predefined region of interest. (b) Depth‐wise analysis identifies physiological and pathological changes in sub‐layers of the cortex (adapted from Northall et al. with permission from Elsevier. Copyright 2023 Elsevier Inc.). (c) Surface‐based analysis maps magnetic susceptibility onto the cortical surface. (d) Network‐based analysis can be defined in two ways: The top panel shows a network derived from independent component analysis (ICA) analysis (adapted from Reeves et al. with permission from Elsevier. Copyright 2022 Elsevier Inc.), while the bottom panel illustrates a network resembling morphological brain networks (adapted from Chen et al. with permission from Springer Nature. Copyright 2024 Springer Nature). (e) Voxel‐wise analysis compares magnetic susceptibility values at each corresponding voxel across all subjects.

Although the aforementioned methods can be effective in certain respects, they often focus on the average effects within specific brain regions. However, pathological changes can occur across the entire brain—not only in the deep gray matter nuclei but also in both the cortex and white matter—making such changes difficult to detect using these approaches. In contrast, voxel‐wise method (Figure 1e) evaluates magnetic susceptibility values at the voxel level, offering an exploratory analysis and identifying changes in spatial distribution. This method avoids the local bias inherent in ROI‐based method by not relying on anatomically predefined regions and provides broader applicability than depth‐wise, surface‐based, or network‐based methods, enabling an unbiased and comprehensive detection. Voxel‐wise method was first applied to study the QSM reconstruction parameters in an AD dataset (Acosta‐Cabronero et al. 2013), followed by a comprehensive processing pipeline (https://gitlab.com/acostaj/QSMexplorer) (Acosta‐Cabronero et al. 2016) and applied to various neurological conditions, including the aging process, PD, WD, and ALS (Acosta‐Cabronero et al. 2016, 2017, 2018; Shribman, Burrows, et al. 2022; Shribman, Bocchetta, et al. 2022). Additionally, derivative processing pipelines have been developed for voxel‐wise analysis (Wang, Martins‐Bach, et al. 2022; Wang et al. 2023).

While extensive literature covers QSM reconstruction algorithms and their clinical applications, systematic reviews focusing specifically on analysis methods, particularly on voxel‐wise analysis, remain scarce. This narrative review therefore aims to address two key issues: (1) the technical details, advantages, and limitations of voxel‐wise whole‐brain analysis for QSM; (2) the clinical applications using voxel‐wise QSM analysis.

For this purpose, the Web of Science (https://www.webofscience.com/wos/) and PubMed (https://pubmed.ncbi.nlm.nih.gov/) databases were surveyed using the keywords “QSM,” “quantitative susceptibility mapping,” “voxel‐wise,” and “voxel‐based” to identify representative studies relevant to voxel‐wise whole‐brain QSM analysis. The search covered publications between 2005 and 2025 and was intended to provide a comprehensive but non‐exhaustive overview of methodological developments and applications in this field. Only original articles and review papers related to brain neuroimaging were included, while books, theses, patents, case reports, conference abstracts, editorials, non‐English publications, and articles not directly relevant to QSM or voxel‐wise analysis were excluded.

2. Processing Pipeline and Methodological Considerations

Voxel‐wise analysis is a powerful technique for exploring individual differences and examining relationships between magnetic susceptibility values and clinical scores at the voxel level. A variety of whole‐brain analysis workflows can be employed, including QSMxT (https://github.com/QSMxT/QSMxT), IronSmith (https://github.com/vzachari/IronSmithQSM), QSMexplorer (https://gitlab.com/acostaj/QSMexplorer), and pipelines within SPM (https://www.fil.ion.ucl.ac.uk/spm/). Their detailed procedures and features are compared in Table S1. The corresponding analysis scripts for these tools have also been organized and can be accessed at https://github.com/tangYING2022/Voxelwise_QSM_Analysis. A typical unified processing pipeline generally comprises three stages: QSM acquisition, QSM reconstruction, voxel‐wise post‐processing (including normalization, smoothing, and statistical analysis). Numerous factors, such as reconstruction algorithms and post‐processing choices, may influence the outcome of voxel‐wise analysis. Key methodological considerations and practical mitigation strategies for voxel‐wise QSM analysis are summarized in Table 1.

TABLE 1.

Key considerations and mitigation strategies for voxel‐wise analysis in QSM.

Section	Issues	Practical mitigation strategy
QSM acquisition	Echo‐time optimization across field strengths (3 T vs. 7 T)	Use field‐strength–specific protocols. 7 T—use shorter TE1 and 3 T—use longer TE1. Use ≥ 3 echoes.
QSM reconstruction	Magnetic susceptibility variation due to differences in reconstruction algorithms	Evaluate reconstruction‐induced variability (e.g., echo combination, inversion algorithms) and conduct sensitivity analysis when feasible.
	Excessive cortical erosion	Refine reconstruction algorithms; use surface‐aware pipelines or study‐specific template; increase resolution.
	Skull base deformations	Optimize background field removal methods; apply field‐map QA; use algorithms robust to large field gradients.
	Large magnetic susceptibility gradients near veins or hemorrhage regions	Apply vessel or high‐magnetic susceptibility masks; consider gradient‐aware reconstruction or regularization tuning.
	Variability in magnetic susceptibility reference choices (e.g., CSF, WM, WB, or no reference)	Align reference selection with study aims and disease characteristics; consider dual‐reference strategies when appropriate.
Voxel‐wise analysis	Using signed or absolute magnetic susceptibility value	Prefer signed magnetic susceptibility value, and perform sensitivity analysis using absolute values.
	Kernel size selection and smoothing and denoising methods	Report the effective smoothing and whether smoothing compensation was performed; consider examining multiple kernel sizes. Use non‐local denoising, edge‐preserving smoothing, or partial‐volume modeling.
	Inconsistent statistical thresholds	Relate thresholds to underlying biological expectations or prior evidence.
Multiple sources	Multiple biological sources and orientation‐dependent magnetic susceptibility in white matter	Apply magnetic susceptibility source separation techniques or joint multi‐contrast modeling to distinguish contributions from different tissue sources.

Abbreviations: CSF: cerebrospinal fluid, QA: quality assurance, TE₁: echo time of the first echo, WB: whole brain, WM: white matter.

2.1. QSM Acquisition

Most studies employ three‐dimensional multi‐echo gradient‐echo (GRE) sequences to enable robust phase evolution modeling and to improve signal‐to‐noise ratio (SNR). The number, spacing, and range of echo times (TEs) jointly determine magnetic susceptibility contrast and noise sensitivity, thereby directly influencing the accuracy and stability of QSM estimates.

Current QSM consensus recommends a short first TE, and a last TE approaching the tissue‐specific T2*, and approximately uniform echo spacing (Bilgic et al. 2024). Importantly, these design principles are field‐strength dependent rather than fixed acquisition rules. With increasing static magnetic field strength, both intrinsic SNR and magnetic susceptibility contrast increase, resulting in higher contrast‐to‐noise ratio (CNR), while the concomitant shortening of T2* enables more time‐efficient sampling of phase evolution. In practice, this leads to typically shorter initial TE and tighter echo spacing at 7 T (e.g., TE₁ ≈ 4 ms, ΔTE ≈ 4 ms), whereas at 3 T, moderately longer values are typically adopted (e.g., TE₁ ≈ 5 ms, ΔTE ≈ 6 ms) to balance phase accrual and signal decay.

2.2. QSM Reconstruction

The reconstruction workflow includes phase unwrapping, multi‐echo data combination, brain mask generation, background field removal, and magnetic dipole field inversion (Figure 2). Phase unwrapping is essential for accurate magnetic susceptibility assessment, as measured phase is constrained within 2π (Robinson et al. 2017). Multi‐echo phase data are typically combined using nonlinear field fitting to obtain the total field, which comprises both tissue‐specific and background fields (Liu et al. 2013). Background fields from magnetic susceptibility sources outside the brain (Schweser et al. 2017), such as air, fat, and bone, can introduce large phase variations that distort the tissue field. Therefore, accurate estimation of tissue magnetic susceptibility requires both exclusion of non‐brain regions via a brain mask (Stewart et al. 2022) and subsequent removal of residual background fields. The final step involves deconvolution of the tissue field with a dipole kernel in k‐space to estimate tissue magnetic susceptibility.

FIGURE 2.

QSM reconstruction pipeline. The dashed boxes indicate the input raw data (phase and magnitude images), while the dashed line indicates an optional component, as some inversion methods may use magnitude images for regularization.

While the procedural details of QSM reconstruction have been extensively summarized in previous reviews, including the recent consensus report by the ISMRM Electro‐Magnetic Tissue Properties Study Group (Bilgic et al. 2024), the present discussion focuses on considerations particularly relevant for voxel‐wise analyses—namely algorithmic combinations, magnetic susceptibility referencing, regional challenges and confounding sources—which critically shape the reliability and interpretability of QSM images.

2.2.1. Combination of QSM Reconstruction Algorithms

Numerous algorithms have been developed for each stage of the QSM reconstruction pipeline, including artificial intelligence (AI)‐integrated approaches. Recent studies have increasingly explored the combination of different reconstruction techniques (Wang, Martins‐Bach, et al. 2022; Santin et al. 2017; Hervouin et al. 2024; Cuna et al. 2023; Naji et al. 2022). For example, the combination of Laplacian boundary value (LBV) method and morphology enabled dipole inversion (MEDI) yields homogeneous QSM images, particularly in the temporal cortex, and demonstrates higher reproducibility than alternative combinations, such as truncated singular value decomposition (TSVD) combined with thresholded k‐space division (TKD), TSVD‐MEDI, and sophisticated harmonic artifact reduction for phase (SHARP)‐MEDI (Santin et al. 2017). Similarly, variable‐SHARP paired with iterative sparse linear equation and least‐squares (iLSQR) results in QSM images with reduced inhomogeneities and greater cross‐subject consistency compared to other background removal methods combined with iLSQR (Wang, Martins‐Bach, et al. 2022). These findings emphasize the importance of considering not only the performance of individual steps but also their interactions. In a systematic evaluation, Salman et al. assessed 126 combinations of background field removal methods (n = 6) and dipole inversion algorithms (n = 21), showing that the choice of background removal method significantly influences reproducibility error and detection sensitivity (Salman et al. 2025). Such outcomes underscore that reconstruction strategies critically shape the ability of voxel‐wise QSM to capture physiological and pathological brain changes.

2.2.2. Magnetic Susceptibility Reference

Because QSM provides relative magnetic susceptibility values, an explicit reference is required to ensure consistency across repeated measurements, scanners, and subjects. The choice of magnetic susceptibility reference remains debated. Several strategies have been proposed, each with strengths and limitations. Cerebrospinal fluid (CSF) has gained popularity for its relative stability but may be unsuitable in the presence of ventricular abnormalities (Tan et al. 2023; Chen, Soldan, et al. 2021; Li, Zhang, et al. 2020). The mean magnetic susceptibility value of the whole brain is less assumption‐dependent but is sensitive to extreme magnetic susceptibility values (Wang et al. 2023; Hagemeier et al. 2018; Zivadinov et al. 2018; Wang, Zhang, et al. 2022). In contrast, omitting explicit referencing avoids biased assumptions about “disease‐free” regions, but increases variability (Acosta‐Cabronero et al. 2017, 2018; Petok et al. 2023). To address these limitations, recent consensus guidelines recommend dual reference strategies, particularly in studies involving diffuse or widespread pathology (Bilgic et al. 2024).

2.2.3. Regional Challenges and Confounding Sources in QSM

Even with optimized algorithmic combinations and proper referencing, magnetic susceptibility estimates remain less reliable in small anatomical structures (e.g., habenular nuclei), in regions affected by skull base distortions (e.g., hippocampus), and in cortical areas prone to biases from blood vessels, lesions, or air–tissue interfaces, where excessive cortical erosion is often required to obtain reliable estimates (Chen, Gong, et al. 2021; Liu et al. 2015). Nonetheless, several preprocessing strategies have been proposed to mitigate these issues in cortical regions, such as surface‐aware pipelines and vessel‐ or blood‐mask strategies (Li, Jacob, et al. 2023; Chen, Soldan, et al. 2025). Moreover, interpretation is further complicated by the mixed sources of magnetic susceptibility signals, including iron, myelin, and other components (Langkammer et al. 2012; Hametner et al. 2018; Stueber et al. 2014; Zheng et al. 2013; Merenstein et al. 2025), as well as orientation dependence with respect to fiber architecture in white matter (Li et al. 2012) and physiological modulations such as tissue oxygenation. Previous studies have predominantly interpreted elevated magnetic susceptibility values as reflecting iron deposition, largely because they focused on deep gray matter, which is known to accumulate iron and exhibit prominent iron‐related magnetic susceptibility changes (Hallgren and Sourander 1958; Haacke et al. 2005; Langkammer et al. 2010). More recently, advances in magnetic susceptibility separation techniques or joint modeling (e.g., χ‐separation, Apart‐QSM, Decomposed‐QSM) (Chen, Gong, et al. 2021; Shin et al. 2021; Li, Feng, et al. 2023) and the integration of complementary physiological measures (e.g., oxygen extraction fraction (OEF) (Engle et al. 2024; Cho et al. 2021) and cerebral metabolic rate of oxygen (CMRO₂) (Zhang et al. 2015)) can help disentangle these contributions, improving both accuracy and interpretability.

2.3. Voxel‐Wise Processing and Statistical Analysis

Voxel‐wise analysis enables whole‐brain assessments of magnetic susceptibility, thereby avoiding the need to predefine anatomical regions of interest. The typical pipeline involves four major stages: spatial normalization, smoothing (with or without smoothing compensation), selection of statistical masks, and voxel‐wise statistical testing (Figure 3). Each of these steps introduces methodological choices that may substantially influence the reproducibility and interpretability of results.

FIGURE 3.

Pipeline for voxel‐wise analysis of QSM. (a) Two types of normalization methods: Direct normalization of native QSM to a QSM template (red area) and indirect normalization using T1‐weighted and magnitude images. Indirect normalization method 1 (purple area) involves a single transformation parameter, while method 2 (blue area) uses two transformation parameters. (b) Detailed steps for smoothing and smoothing compensation. In this panel, * denotes convolution operation, and $÷$ denotes division operation. (c) Different masks used for statistical analysis: The blue mask (top) covers the brain parenchyma, the green mask (middle) covers gray matter, and the pink mask (bottom) represents the substantia nigra.

2.3.1. Normalization

Normalization is essential in voxel‐wise analyses, allowing inter‐subject comparison by aligning individual QSM images into a standardized space. It accounts for variations in intracranial volume, subject positioning, and scanning geometry. Two main approaches are employed (Figure 3a): direct normalization of QSM images to a template (red area in Figure 3a) and indirect normalization via structural MRI (purple and blue areas in Figure 3a).

2.3.1.1. Direct Normalization

In direct normalization, native QSM images are directly normalized to a QSM template. This approach is computationally straightforward and has been more frequently applied in ROI‐wise analyses (Wu et al. 2023; Yan et al. 2023; Jiang et al. 2023). In voxel‐wise studies, direct normalization may be considered when high‐quality T1‐weighted images (T1‐WI) are unavailable or cannot be relied on for cross‐modality normalization with magnitude images. However, anatomical correspondence in QSM images can be compromised by magnetic susceptibility‐dependent contrast and background field removal steps introduced during QSM reconstruction. These factors may reduce the fidelity of anatomical landmarks and consequently limit normalization accuracy. As a result, the applicability of direct normalization for precise voxel‐wise inference remains under active investigation.

2.3.1.2. Indirect Normalization

Indirect normalization maps QSM images to a template space via intermediate structural images, thereby reducing the direct influence of magnetic susceptibility‐dependent contrast on spatial normalization. According to the number of transformation parameters applied to QSM data, this strategy can be further categorized into single‐parameter and dual‐parameters normalization. In the single‐parameter normalization (purple panel in Figure 3a), QSM images are normalized using a single transformation matrix. In the common implementation, native T1‐WI are first normalized to the first‐echo magnitude images, followed by aligning the normalized T1‐WI to a template space. The resulting transformation matrix (1st normalization matrix in Figure 3a) can be directly applied to QSM images (Nepozitek et al. 2023; Chen et al. 2022). When T1‐WI and QSM images are derived from the same acquisition or inherently share the same spatial geometry, this procedure can be simplified by applying the T1‐to‐template transformation to QSM images. In these cases, QSM data undergo only one spatial transformation, thereby minimizing interpolation‐induced deformation and reducing potential bias (Varga et al. 2024). In dual‐parameters normalization (blue panel in Figure 3a), QSM images are normalized using two concatenated transformation matrices. Specifically, the magnitude images are first normalized to native T1‐WI, which are then aligned to a T1‐weighted template (2nd and 3rd normalization matrices in Figure 3a). These two matrices are combined to normalize QSM images into a unified space (Shribman, Bocchetta, et al. 2022; Wang et al. 2023; Ravanfar et al. 2023).

From a practical perspective, the choice between single‐ and dual‐parameters normalization should align with study priorities. Single‐parameter normalization is generally preferred when preservation of native QSM spatial fidelity is prioritized, for example in high‐resolution datasets or studies focusing on fine‐scale magnetic susceptibility variations. In contrast, dual‐parameters normalization is more suitable when robustness is emphasized, particularly in multi‐site studies, heterogeneous clinical populations, or datasets with variable magnetic susceptibility contrast and prominent artifacts when selecting the normalization strategy. A decision workflow to guide this selection is provided in Figure S1b.

2.3.1.3. Template Selection

Beyond the choice of normalization strategy, the selection of an appropriate template space represents another important factor influencing spatial accuracy and potential bias in voxel‐wise QSM analyses. Most studies have used the MNI152 T1 template (Grabner et al. 2006). However, population differences (e.g., disease‐specific atrophy patterns) and site‐related variability in multi‐center datasets may introduce systematic bias. To address this, study‐specific templates that better capture the structural features of the target cohort are often recommended (Keller et al. 2004). Figure S1a outlines a practical decision tree to inform this key choice.

2.3.2. Smoothing

2.3.2.1. Implementation

Smoothing is commonly applied to QSM images after spatial normalization to increase the SNR and to facilitate valid statistical inference (e.g., random field theory, RFT) by better satisfying underlying assumptions and reducing inter‐subject anatomical variability (Soares et al. 2016). Unlike other MRI modalities, QSM introduces smoothing during reconstruction through filtering, which should be considered when selecting post‐normalization Gaussian kernels. Depending on the study design and characteristics of the data, smoothing may be performed alone or in combination with smoothing compensation to mitigate partial volume effects (Lee et al. 2009). Typically, smoothing compensation involves convolving the QSM images with a three‐dimensional Gaussian kernel, which is also applied to the corresponding brain mask, followed by dividing the smoothed QSM images by the smoothed mask to generate smoothed‐compensated images (Figure 3b).

2.3.2.2. Kernel Selection

Gaussian kernels are commonly reported in the range of 2–8 mm full width at half maximum (FWHM), corresponding approximately to a Gaussian kernel with a standard deviation (σ) of 2–3.5 mm (Shribman, Bocchetta, et al. 2022; Nepozitek et al. 2023; Chen et al. 2022; Uchida et al. 2019), and the optimal kernel size depends on native voxel resolution, expected anatomical scale, and the statistical inference strategy. In practice, a minimum kernel size of approximately twice the voxel dimensions is often recommended to satisfy statistical assumptions for RFT‐based inference (Mikl et al. 2008), while studies targeting small anatomical structures may require a smaller kernel to preserve spatial specificity. To better reflect the spatial autocorrelation relevant to RFT‐based inference and to mitigate normalization errors, we recommend reporting effective smoothing, defined as the convolution of reconstruction filtering with the post‐normalization Gaussian smoothing kernel. Caution is advised when applying isotropic smoothing kernels to images with anisotropic voxel dimensions (e.g., 1.0 × 1.0 × 2.0 mm³), as this can lead to directional bias and misinterpretation of spatial extent; appropriate adjustment of kernel dimensions or resampling to isotropic voxels prior to smoothing is recommended.

2.3.2.3. Alternative Strategies

Although Gaussian smoothing improves SNR, it may introduce blurring, spatial bias, and dependency on voxel size. To mitigate noise while preserving fine structural details, several alternative strategies have been proposed, including non‐local denoising, edge‐preserving filters, and partial‐volume modeling. Non‐local denoising preserves complex textures and fine structural details (Manzano Patron et al. 2024; Mishro et al. 2022), while edge‐preserving smoothing reduces noise without blurring critical gray‐white matter interfaces (Zhu et al. 2019). Partial‐volume modeling further refines this by estimating tissue composition within each voxel to mitigate signal mixing at tissue boundaries (Henf et al. 2018).

2.3.2.4. Considerations and Future Directions

A potential consideration is that these nonlinear and spatially adaptive filters may introduce spatial non‐stationarity into the noise field, which could theoretically challenge the stationarity assumption required by RFT and complicate parametric cluster‐level inferences. To address this, voxel‐wise QSM studies employing such strategies may benefit from statistical approaches that are robust to non‐stationarity, such as non‐parametric permutation‐based inference, thereby improving the reliability of whole‐brain conclusions. Although these advanced denoising techniques have not yet been applied in voxel‐wise QSM analysis, they hold potential for improving SNR and preserving fine structural details, and future studies should evaluate their performance and compatibility with various statistical frameworks.

2.3.3. Statistical Analysis

Voxel‐wise QSM analyses require careful statistical considerations regarding the choice of statistical mask, magnetic susceptibility measure, statistical inference approach, and thresholding/reporting criteria.

2.3.3.1. Statistical Mask Selection

As illustrated in Figure 3c, the use of statistical masks serves to restrict statistical testing to brain regions of interest and reduce the burden of multiple comparisons. Masks can be designed to target specific regions, structures, or pathologies, depending on the research question. For example, a gray matter mask was used to exclude myelin‐related effects in a study on multiple sclerosis (MS) (Pontillo et al. 2023), while masks targeting voxel‐level changes in deep gray nuclei and limbic structures have been employed in AD research (Kuchcinski et al. 2023). In contrast, Acosta‐Cabronero et al. extended their analysis in ALS to encompass the entire brain (Acosta‐Cabronero et al. 2018). Compared to the ROI‐based method, voxel‐wise analysis offers more detailed information on the spatial distribution of magnetic susceptibility changes across broader brain regions. The decision between region‐specific and whole‐brain masks should be guided by both biological plausibility and statistical considerations. Region‐focused masks may improve power for detecting localized effects, while whole‐brain masks support exploratory analyses and the identification of unexpected spatial patterns.

2.3.3.2. Signed Versus Absolute Magnetic Susceptibility Values

A key decision is whether to use signed or absolute magnetic susceptibility values for statistical analysis. Signed values preserve polarity and thus retain biological interpretability, enabling differentiation between paramagnetic sources (e.g., iron) and diamagnetic substances (e.g., myelin, or calcium) (Ravanfar et al. 2023, 2022; Kuchcinski et al. 2023; Uchida, Kan, Sakurai, Arai, et al. 2020; van Bergen et al. 2018). In contrast, absolute values suppress polarity by construction and can increase statistical efficiency by mitigating blooming artifacts and signal cancelation, often yielding larger effect sizes at the cost of biological specificity (Shribman, Bocchetta, et al. 2022; Betts et al. 2016; Yang et al. 2022; Thomas et al. 2020). This distinction is critical in conditions where oppositely signed magnetic susceptibility sources coexist within the same anatomical region. Representative examples include MS, where iron‐rich active rims surround demyelinated lesion cores, and hereditary cerebral hemorrhage with amyloidosis–Dutch type, in which iron deposits and calcification jointly contribute to the characteristic ‘cortical stripes’ pattern (Hametner et al. 2013; Bulk et al. 2018). In such scenarios, absolute values would conflate distinct pathological processes, whereas signed values preserve their polarity‐dependent signatures. Accordingly, we recommend using signed magnetic susceptibility values for primary biological inference with polarity‐aware interpretation, while reserving absolute values for sensitivity or complementary analyses. Established source separation techniques may further enable the disentanglement of mixed magnetic susceptibility sources at the voxel level.

2.3.3.3. Inference and Multi‐Comparisons Correction

Voxel‐wise analyses in QSM typically employ nonparametric permutation tests with threshold‐free cluster enhancement (TFCE) or RFT‐based inference. Permutation tests provide a robust, distribution‐free approach for significance testing, while TFCE enhances detection power by integrating signal intensity and spatial extent (Smith and Nichols 2009). When applied together, permutation‐based TFCE enables sensitive yet rigorous inference. In contrast, RFT offers a parametric framework by modeling the statistical map as a Gaussian random field, thereby relying on assumptions regarding its smoothness and distribution (Mikl et al. 2008). Among methods for multiple comparison correction, family‐wise error (FWE) correction across all voxels is the most rigorous for controlling false positives, while false discovery rate (FDR) correction is also widely adopted as a common alternative. Software tools such as SPM and FSL (https://fsl.fmrib.ox.ac.uk/fsl/docs/#/) are commonly employed to perform these statistical analyses.

2.3.3.4. Thresholding and Reporting Standards

Variability in significant thresholds across QSM studies complicates the comparison of results. The thresholds of p _FWE < 0.05 and p _FDR < 0.05 are commonly used. In addition, some studies report uncorrected results (p < 0.001) or employ heterogeneous cluster‐size criteria (10–100 voxels) (Li, Zhang, et al. 2020; Petok et al. 2023; Chen et al. 2022; Uchida et al. 2019; Pontillo et al. 2023; Duan et al. 2022; Margoni et al. 2023; Xu et al. 2023; Adams et al. 2023; Uchida, Kan, Sakurai, Inui, et al. 2020; Kano et al. 2023). These discrepancies could account for the variability in reported findings.

2.3.4. Recommendations and Caveats

2.3.4.1. Implementation Recommendations

To facilitate reproducible implementation and standardized reporting, a minimum analysis and reporting standard is proposed (summarized in Table 2), together with a concise reporting checklist for voxel‐wise QSM studies provided in Box S1. For data acquisition and QSM reconstruction, alignment with existing QSM consensus recommendations is encouraged (Bilgic et al. 2024). Nevertheless, it is critical to visually inspect both the field maps and the final QSM images in each study to verify the absence of severe artifacts and ensure reconstruction quality. In preprocessing, detailed reporting of the template, normalization pipeline, smoothing procedures is recommended. Following these steps, signed magnetic susceptibility values may serve as the primary metric, with absolute magnetic susceptibility values available for sensitivity analyses. For statistical inference, non‐parametric permutation testing integrated with TFCE is advised. Results should be evaluated against a significance threshold (e.g., p _FWE < 0.05 or p _FDR < 0.05) or alternatively a cluster‐based criterion (voxel‐level p < 0.001, cluster‐level p _FWE < 0.05, with a minimum extent of 50 voxels).

TABLE 2.

Items to report and recommendations in the voxel‐wise QSM analysis.

Section	Items	Specific recommendation to report
Methods	Study population	Sample size, age, sex; inclusion/exclusion criteria; site/batch information
	MRI acquisition and preprocessing	Scanner type, sequence; QSM reconstruction; template name/version with voxel‐size; normalization; transforms applied to which images; interpolation method/order; smoothing, effective smoothing kernel; modulation/ICV handling; statistical mask
	Statistical model	Signed or absolute magnetic susceptibility; GLM or equivalent; covariates: age, sex, ICV, motion/QA metrics, site/batch
	Inference and multiple comparison correction	RFT or Permutation with TFCE; specify correction type (FWE/FDR/uncorrected); cluster‐defining threshold if applicable; report voxel‐wise and cluster‐wise results; share full unthresholded maps
Results	Primary findings	Significant clusters: location, size, peak coordinates; effect sizes if available
	Voxel‐wise maps	Thresholded voxel‐wise images; share full unthresholded maps for reproducibility
	Supplementary analyses	Sensitivity analyses (thresholds, covariate models); subgroup analyses if relevant

Abbreviations: FDR: false discovery rate, FWE: family‐wise error, GLM: general linear model, ICV: intracranial volume, QA: quality assurance, RFT: random field theory, TFCE: threshold‐free cluster enhancement.

For result presentation, significant clusters should be summarized in tables (including location, size, and neuroanatomical region) and visualized on appropriate brain templates. To further support transparency, we encourage sharing of complete unthresholded statistical maps as Supporting Information or via a public repository.

2.3.4.2. Confounds and Limitations

Beyond these technical aspects, voxel‐wise analyses also face broader methodological challenges. Voxel‐wise analyses require well‐defined hypotheses, as inappropriate statistical assumptions may yield spurious results. Correction for multiple comparisons, while essential to control false positives, can be overly conservative and obscure subtle changes in cortical layers or subregions (Merenstein et al. 2024; Deistung et al. 2013; Lee et al. 2023). Additionally, the group‐level nature of voxel‐wise analyses may mask clinically relevant individual variations, as subtle subject‐specific effects can be lost in the group average (Essex et al. 2025). These limitations highlight the need for rigorous hypothesis design, harmonized statistical practices, and integration with multimodal or multiscale approaches, which may improve both reproducibility and biological interpretability in voxel‐wise QSM studies.

3. Applications of Voxel‐Wise Whole‐Brain QSM Analysis

Voxel‐wise whole‐brain QSM analysis is a powerful tool for investigating various neurological and neuropsychiatric disorders. This section explores its applications in studying aging, genetic‐phenotype (Figure 4), neurodegenerative diseases including AD, PD, MS, ALS, and WD (Figure 5), neuropsychiatric disorders such as alcohol use disorder (AUD) and attention‐deficit/hyperactivity disorder (ADHD), and other conditions (Figure 6). Tables S2–S4 provide a structured summary of the literature: Table S2 summarizes MRI acquisition parameters and QSM reconstruction methods, Table S3 synthesizes cohort characteristics, voxel‐wise processing, and key outcomes, and Table S4 consolidates consistent magnetic susceptibility findings across different disorders.

FIGURE 4.

Voxel‐wise QSM applications in aging and genetic‐phenotype. (a) Regression analysis between age and magnetic susceptibility, with results standardized to MNI space. Copyright 2016, Acosta‐Cabronero et al. Published by the Journal of Neuroscience. (b) Voxel‐wise associations between magnetic susceptibility and phenotypic traits, including tea consumption and frequency of alcohol intake (six or more units). Copyright 2022, Wang et al. Published by Springer Nature. p _FWE; p value with family‐wise error correction.

FIGURE 5.

Applications of voxel‐wise QSM analysis in neurodegenerative disorders. (a) Regression analysis of absolute magnetic susceptibility values versus standardized uptake value ratio (tau‐SUVR) across the whole gray matter in Alzheimer's disease, mapped to MNI coordinates. Copyright 2020, Spotorno et al. Published by Oxford University Press on behalf of the Guarantors of Brain. (b) Elevated absolute magnetic susceptibility values in patients with Parkinson's disease versus healthy controls, results aligned to MNI coordinates. Copyright 2020, Thomas et al. Published by BMJ Publishing Group Ltd. (c) Increased absolute magnetic susceptibility values in neurological Wilson's disease (WD) patients compared to those with hepatic WD, showing a positive correlation between magnetic susceptibility values and Unified Wilson's Disease Rating Scale‐Neurological (UWDRS‐N) scores. Copyright 2021, Shribman et al. Published by Oxford University Press on behalf of the Guarantors of Brain. L: left; p _FWE; p value with family‐wise error correction; R: right.

FIGURE 6.

Voxel‐wise QSM applications beyond neurodegenerative diseases. (a) Voxel‐wise QSM analysis of normal‐appearing white matter in patients with multiple sclerosis (MS) versus healthy controls (HC). Copyright 2022, Reza Rahmanzadeh. Published by Elsevier Inc. (b) Voxel‐wise QSM comparison between patients with cerebral small vessel disease (CSVD) and healthy controls (HC). Copyright 2025, Wang et al. Published by Alzheimer's & Dementia, Wiley Periodicals LLC on behalf of the Alzheimer's Association. (c) Voxel‐wise comparison among patients in stage 5 non‐dialysis chronic kidney disease (CKD 5ND), those with stages 1–4, and healthy controls. Copyright 2024, Song et al. Published by CNS Neuroscience & Therapeutics, John Wiley & Sons Ltd. p _FDR: p value with false discovery rate.

3.1. Aging and Genetic‐Phenotype

3.1.1. Aging

Iron deposition is a hallmark of aging, contributing to neuronal cell death and axonal degeneration (Kurz et al. 2008). Early anatomical studies have identified iron deposition in deep gray matter and cortex during aging (Hallgren and Sourander 1958; Spatz 1922; Bartzokis et al. 2007). Voxel‐wise QSM analysis allows detailed in vivo mapping of these changes. Acosta‐Cabronero et al. analyzed 116 healthy adults aged 20–79 years (Acosta‐Cabronero et al. 2016), revealing increased magnetic susceptibility in elder adults in the striatum, midbrain, and motor, premotor, posterior insular, superior prefrontal, and cerebellar cortices, indicating iron accumulation in these regions (Figure 4a). These observations were corroborated by Burgetova et al. (2021).

Given the multifactorial nature of aging (Higgins‐Chen et al. 2021), integrating voxel‐wise QSM with multimodal imaging, genetics, and other relevant factors is crucial for comprehensive assessment of aging‐related changes.

3.1.2. Phenotypic and Genetic Factors

Understanding the relationship between magnetic susceptibility changes and phenotypic as well as genetic factors is essential for elucidating the biological mechanisms underlying aging and disease. Voxel‐wise correlation analyses in 35,273 UK Biobank participants revealed both positive and negative associations between magnetic susceptibility and various phenotypic factors, such as anemia, diabetes, tea intake, and alcohol consumption, spanning subcortical, cerebellar, and several white matter regions (part of which are shown in Figure 4b) (Wang, Martins‐Bach, et al. 2022). Additionally, genes related to iron transport, calcium homeostasis, and glial/myelin function were linked to magnetic susceptibility variations in the caudate, pallidum, substantia nigra, thalamus, red nucleus, cerebellar nuclei, and white matter. Notably, some genes influenced magnetic susceptibility indirectly, suggesting complex genetic regulation.

3.2. Neurodegenerative Diseases

3.2.1. Alzheimer's Disease

AD has been a central focus of QSM research among neurodegenerative diseases. Accumulation of amyloid‐beta (Aβ) plaques (defining AD pathology) and tau neurofibrillary tangles (characterizing AD disease burden) is closely associated with iron deposition (Sayre et al. 2000; Everett et al. 2014; Roberts et al. 2017; Zeineh et al. 2015), with cognitive impairment linked to microglial iron‐driven neuroinflammation and neuronal ferroptosis (Ayton et al. 2020; Ward et al. 2014; Tran et al. 2022).

Increased magnetic susceptibility in AD patients has been observed in the cerebral and cerebellar cortices and subcortical nucleus (including caudate, hippocampus, and amygdala) (Yang et al. 2022; Kim et al. 2017), while MCI—the prodromal phase of AD—shows no significant difference in magnetic susceptibility compared to healthy controls (Li, Fan, et al. 2024). AD subtypes exhibit distinct patterns: patients with limbic‐predominant AD exhibited the highest magnetic susceptibility value in the hippocampal head, while those with hippocampal‐sparing AD showed higher magnetic susceptibility in the striatum and thalamus (Kuchcinski et al. 2023).

The interplay between iron burden, proteinopathy (Aβ and tau accumulation), and cognitive decline has become a major focus of ongoing investigation. Voxel‐wise correlation analyses revealed that regional magnetic susceptibility positively correlates with Aβ in frontal/temporal cortices and basal ganglia (van Bergen et al. 2018), and with tau across temporal, parietal, precuneus, cingulate, occipital, and frontal regions (Figure 5a) (Spotorno et al. 2020). Additionally, higher magnetic susceptibility in the right parietal and lateral occipital cortices is associated with lower cognitive performance on Mini‐Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) (Yang et al. 2022). While ROI‐based studies have linked cognitive decline—assessed through both global composite scores and domain‐specific measures—and magnetic susceptibility (Chen, Soldan, et al. 2021; van Bergen et al. 2018), voxel‐wise analyses specifically examining these relationships across the entire brain remain limited.

Nevertheless, correlations between magnetic susceptibility and Aβ or tau biomarkers remain weak or inconsistent across studies. This dissociation likely stems from differences in disease stage, cohort characteristics, and limitations of QSM or positron emission tomography (PET) and CSF assays. Along the AD continuum, Aβ and tau pathologies follow distinct temporal trajectories, with Aβ accumulation occurring early and plateauing before symptoms appear (Jack Jr. et al. 2010; Villemagne et al. 2013), while magnetic susceptibility changes worsen in later disease stages. In cross‐sectional studies spanning preclinical, MCI, and dementia phases, these processes may be dissociated, leading to weak or absent correlations. Additionally, magnetic susceptibility measurements are influenced by factors beyond core AD pathology, including age‐related iron accumulation, small vessel disease, and microbleeds, and are further complicated by QSM's inability to distinguish the chemical forms of iron or its uniform sensitivity to different iron states. Moreover, the relatively low spatial resolution of PET imaging and off‐target binding of tau tracers further limit the detection of robust voxel‐wise or regional correlations (Uchida et al. 2022).

Beyond investigations of AD stages, subtypes, and clinical correlations, therapeutic developments—such as anti‐Aβ monoclonal antibodies (e.g., lecanemab and donanemab)—have demonstrated efficacy in managing early‐stage cognitive decline (Neațu et al. 2023; Behl 2024; Smith and Ownby 2024). Although longitudinal QSM studies suggest that magnetic susceptibility changes associated with disease progression often require longer follow‐up periods, likely due to the net contribution of multiple competing paramagnetic and diamagnetic tissue components, emerging evidence indicates that QSM may still be informative over shorter time frames in treatment contexts. For example, a recent study in early AD patients receiving biweekly intravenous lecanemab demonstrated a transient increase in the blood–brain barrier (BBB) water exchange rate prior to the onset of amyloid‐related imaging abnormalities with microhemorrhages or superficial siderosis (ARIA‐H), accompanied by a sustained elevation in paramagnetic magnetic susceptibility (Uchida et al. 2025). These findings suggest that paramagnetic magnetic susceptibility may be sensitive to treatment‐related pathological changes and could serve as a useful imaging marker for monitoring therapeutic effects and safety.

3.2.2. Parkinson's Disease

PD, the second most prevalent neurodegenerative disease after AD, is preceded by isolated rapid eye movement sleep behavior disorder (iRBD) and manifests as tremors, bradykinesia, and muscular rigidity (Bloem et al. 2021). Its pathophysiology involves loss of dopaminergic neurons in the substantia nigra (SN) with α‐synuclein deposition (Fearnley and Lees 1991; Kordower et al. 2013; Damier et al. 1999), and iron accumulation has been proposed as a potential mechanism (Morris et al. 2018).

Synthesis of the voxel‐wise QSM literature reveals a consistent pattern of increased magnetic susceptibility in the SN of patients with PD. Among the five identified studies employing voxel‐wise analysis (all conducted at 3 T), four (80%) reported significantly elevated magnetic susceptibility in the SN or its functional subregions (Acosta‐Cabronero et al. 2017; Varga et al. 2024; Du et al. 2016; Wen et al. 2025). The sole exception applied an analytical mask that excluded the SN; however, its subsequent targeted ROI‐based analysis demonstrated increased magnetic susceptibility within the SN, thereby aligning with the overall pattern once this methodological limitation was addressed (Thomas et al. 2020). Beyond the midbrain, two voxel‐wise QSM studies that extended whole‐brain analyses to cortical regions (both at 3 T) reported aberrant magnetic susceptibility patterns predominantly involving the temporal and prefrontal cortices (Figure 5b) (Acosta‐Cabronero et al. 2017; Thomas et al. 2020). Increased magnetic susceptibility has also been observed in subcortical structures, including the striatum and rostral pons (Acosta‐Cabronero et al. 2017; Thomas et al. 2020).

In patients with iRBD, elevated magnetic susceptibility has been reported in the frontal–parietal cortex and ventromedial SN compared with healthy controls. Relative to individuals with iRBD, patients with PD exhibit higher magnetic susceptibility in the hippocampus, amygdala, and pars compacta of SN, alongside lower magnetic susceptibility in frontal and parietal cortices (Varga et al. 2024). Furthermore, PD patients with MCI show higher magnetic susceptibility than cognitively normal PD patients and healthy controls, particularly in the caudate head and putamen (Uchida et al. 2019; Uchida, Kan, Sakurai, Inui, et al. 2020).

Additionally, magnetic susceptibility correlates with both motor and cognitive outcomes. For motor symptoms, higher Unified Parkinson's Disease Rating Scale part III (UPDRS‐III) scores correlate with increased magnetic susceptibility in the putamen (Thomas et al. 2020), and longitudinal studies link motor severity with increased magnetic susceptibility in the insular cortex, basal ganglia, SN, red nucleus, and dentate nucleus (Thomas et al. 2024). The relationship between magnetic susceptibility and cognitive performance in PD remains heterogeneous. While some cross‐sectional studies report negative correlations with MoCA scores in the hippocampus, thalamus, caudate, cuneus, amygdala, and fusiform gyrus (Uchida et al. 2019; Thomas et al. 2020), longitudinal evidence further implicates additional regions—temporal cortex, putamen, basal forebrain, brainstem, and insula over time (Thomas et al. 2024). However, one study found no such association (Varga et al. 2024). These discrepancies likely reflect differences in cohort composition and analysis pipeline configuration. In particular, studies focusing on early‐stage PD with limited cognitive variance may have reduced sensitivity to detect magnetic susceptibility–cognition relationships, an effect that can be further amplified by methodological choices such as small voxel sizes and minimal spatial smoothing, which lower SNR in voxel‐wise analyses.

Despite advances, early identification and accurate diagnosis remain challenges (Tolosa et al. 2021). Future work should leverage voxel‐wise QSM‐derived features to delineate PD subtypes, predict progression, and enable precision therapy.

3.2.3. Amyotrophic Lateral Sclerosis

ALS is a relatively rare central nervous system disease characterized by progressive upper and lower motor neuron dysfunction and muscle weakness (Feldman et al. 2022). A hallmark of ALS pathology is the presence of transactive response DNA‐binding protein 43 kDa (TDP‐43) proteinopathy (Neumann et al. 2006).

Accumulating evidence suggests that iron dysregulation plays an important role in ALS pathogenesis (Wang and Liu 2015; Jeong et al. 2009). Whole‐brain QSM studies have reported elevated magnetic susceptibility in the motor cortex, substantia nigra, globus pallidus, and red nucleus, alongside reduced magnetic susceptibility in the corticospinal tract in ALS compared with controls (Acosta‐Cabronero et al. 2018). These findings are consistent with postmortem observations, including TDP‐43 immunoreactivity, the loss of Betz cells, and the presence of ferritin‐positive microglia and macrophages (Brettschneider et al. 2013; Udaka et al. 1986; Kwan et al. 2012).

Current challenges in ALS include early diagnosis, clarification of disease mechanisms, and prediction of progression (González‐Sánchez et al. 2025). QSM holds distinct promise for noninvasively quantifying iron deposition, closely linked to ALS pathogenesis. Voxel‐wise and longitudinal analyses may facilitate the identification of early iron‐related neurotoxicity biomarkers, enabling earlier detection, more accurate monitoring, and improved prognostic evaluation in ALS.

3.2.4. Wilson's Disease

WD, an autosomal‐recessive disorder of copper metabolism, is caused by mutations in the P‐type adenosine triphosphatase (ATP7B) gene (Roberts and Schilsky 2023). Clinically, WD can be categorized into neurological and hepatic types, or into active and stable types, based on clinical symptoms and neurological status. Pathological features include copper deposition, frequently accompanied by iron deposition and demyelination (Meenakshi‐Sundaram et al. 2008).

These pathological changes can be quantitatively assessed using unbiased whole‐brain QSM analysis. Higher magnetic susceptibility has been observed in patients with WD compared to healthy controls, especially in the caudate nucleus, putamen, globus pallidus, SN, and red nucleus (Fan et al. 2024). Furthermore, patients with neurological WD exhibit higher magnetic susceptibility in the globus pallidus and putamen than those with hepatic WD (Fan et al. 2024). Elevated values have also been observed in both gray matter regions (cingulate and medial frontal cortices) and white matter regions (internal and external capsules, corpus callosum, and corona radiata) in neurological WD patients (Figure 5c) (Shribman, Bocchetta, et al. 2022).

Correlations between magnetic susceptibility and neurological severity, neuropsychological test scores, and plasma markers have been reported (Shribman, Burrows, et al. 2022; Shribman, Bocchetta, et al. 2022; Fan et al. 2025), with stable‐type WD showing positive correlations between Unified Wilson's Disease Rating Scale (UWDRS‐N) scores and magnetic susceptibility in widespread cortical and cerebellar regions (Shribman, Bocchetta, et al. 2022). However, no significant associations were found between neuropsychological test scores and magnetic susceptibility in WD patients (Shribman, Burrows, et al. 2022). Furthermore, in patients untreated with chelators, the positive association between plasma glial fibrillary acidic protein (GFAP) levels and magnetic susceptibility in the putamen and cuneus gyrus reflects a correlation between iron deposition and neuroinflammation (Fan et al. 2025).

3.3. Neuropsychiatric Disorders

3.3.1. Alcohol Use Disorder

AUD, a chronic relapsing condition, is among the most prevalent addiction‐related disorders globally. Voxel‐wise QSM study implicates iron deposition in its pathophysiology (Tan et al. 2023) with 186 AUD patients and 274 healthy controls. This study revealed increased magnetic susceptibility in the dorsal striatum (putamen and caudate), with positive correlations to obsessive‐compulsive drinking scale scores. The authors attributed this increased magnetic susceptibility to alcohol‐induced enhancement of intestinal iron absorption and compromised blood–brain barrier (BBB) integrity.

Further exploring the mechanisms of BBB disruption in AUD, Adams et al. found alcohol‐related enhancement of intestinal iron absorption in AUD patients and higher magnetic susceptibility in the left globus pallidus (Adams et al. 2023). However, no direct correlation was observed between serum ferritin levels and total brain iron, suggesting that brain iron homeostasis is regulated by the BBB and iron transport proteins, independent of systemic iron levels.

Current pharmacotherapies, including disulfiram, acamprosate, and naltrexone, show variable efficacy (Witkiewitz et al. 2019), complicated by the high comorbidity of AUD and depression (McHugh and Weiss 2019). QSM holds potential for stratifying patients based on neurobiological markers of iron dysregulation, facilitating individualized treatment strategies and the development of clinically relevant biomarkers.

3.3.2. Attention‐Deficit/Hyperactivity Disorder

In addition to AUD, whole‐brain QSM analysis has been applied to study other mental illnesses, such as ADHD. ADHD is a neurodevelopmental disorder typically characterized by impulsivity, age‐inappropriate inattention, and hyperactivity (May et al. 2023). To investigate the role of iron in children with ADHD, Chen et al. performed a voxel‐wise QSM analysis (Chen et al. 2022). Their findings revealed decreased magnetic susceptibility in several regions in ADHD patients, including the bilateral striatum, anterior cingulum, olfactory gyrus, and right lingual gyrus. These findings are notable given the established role of iron‐related processes in neurotransmitter synthesis and dopaminergic function, which are central to ADHD pathophysiology.

Building on these whole‐brain findings, it is important to note that iron is essential for normal neurodevelopment. In typically developing children, brain magnetic susceptibility shows a gradual increase with age (Treit et al. 2021; Li, Tong, et al. 2023), reflecting ongoing developmental processes. Using combined QSM and BBB imaging, Uchida et al. demonstrated that age‐dependent magnetic susceptibility changes are closely associated with BBB water exchange during childhood, suggesting coordinated maturation of iron‐related magnetic susceptibility and BBB function (Rocca et al. 2010). In ADHD, the observed reductions in magnetic susceptibility raise the possibility that altered BBB function or dysregulated iron‐related processes contribute to atypical brain development. Future studies integrating voxel‐wise QSM with BBB permeability measures may help clarify the mechanisms underlying magnetic susceptibility alterations in ADHD and improve the clinical interpretability of QSM in neurodevelopmental disorders.

3.4. Others

3.4.1. Multiple Sclerosis

MS is a prevalent demyelinating disorder. Its pathology involves a complex interplay of pathological processes, including iron deposition, axonal loss, demyelination, and inflammation (Filippi et al. 2018). Voxel‐wise QSM studies have consistently reported increased magnetic susceptibility in the caudate, putamen, globus pallidus, and thalamus of MS patients (Zivadinov et al. 2018; Cobzas et al. 2015).

Subtype‐specific differences in magnetic susceptibility patterns have also been observed: in secondary progressive MS (SPMS), magnetic susceptibility is reduced in the thalamus, particularly in the left pulvinar, but elevated in the dorsal globus pallidus compared to those with relapsing–remitting MS (RRMS) (Zivadinov et al. 2018). Detailed thalamic analyses further show decreased magnetic susceptibility across most thalamic subregions (excluding the pulvinar nucleus) in RRMS patients relative to matched controls, while SPMS patients displayed more pronounced reductions specifically in the pulvinar nucleus (Zivadinov et al. 2018). These findings underscore the vulnerability of deep gray matter, especially the thalamus, to iron dysregulation in MS.

Altered magnetic susceptibility has been linked to cognitive dysfunction in MS patients (Rocca et al. 2010), highlighting the clinical relevance of magnetic susceptibility measurements in MS research. Beyond gray matter, group differences have also been detected in normal‐appearing white matter, where both increased and decreased magnetic susceptibility have been observed compared with controls (Figure 6a).

3.4.2. Cerebral Small Vessel Disease

Cerebral small vessel disease (CSVD) is a chronic cerebrovascular condition affecting small arteries, arterioles, capillaries, and veins (Pantoni 2010). Iron deposition in CSVD has been implicated in oxidative phosphorylation, myelination, and neurotransmitter metabolism (Ward et al. 2014). Voxel‐wise QSM study has shown significantly increased magnetic susceptibility in the bilateral deep gray matter (Figure 6b), particularly in the putamen, caudate, and substantia nigra in CSVD patients compared with controls (Wang et al. 2025), whereas brainstem regions exhibited relative reductions.

Cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy (CADASIL), a rare hereditary CSVD caused by NOTCH3 mutations (METACOHORTS Consortium 2016), also demonstrates abnormal iron deposition (Liem et al. 2012; Sun et al. 2020). Voxel‐wise analysis revealed increased magnetic susceptibility in deep gray matter regions (caudate, putamen, thalamus) and cortical areas (middle temporal gyrus and bilateral precuneus extending into the lateral occipital gyrus) (Jia et al. 2023). These findings align with observations reported in both symptomatic and asymptomatic patients carrying gene mutations. Additionally, symptomatic patients exhibit higher magnetic susceptibility than asymptomatic carriers in the putamen, pallidum, caudate nucleus, temporal pole, and centrum semiovale, whereas asymptomatic carriers show localized increases in the putamen, caudate nucleus, temporal pole, and centrum semiovale (Uchida, Kan, Sakurai, Arai, et al. 2020). These alterations may reflect secondary neurodegeneration and be linked to increased BBB permeability.

Aligned with international consensus guidelines for neuroimaging standards in CSVD research (Duering et al. 2023), future studies should leverage voxel‐wise QSM to: (1) delineate mechanisms underlying periventricular versus deep white matter hyperintensities; (2) assess cerebral microbleed distribution for diagnosis; and (3) establish QSM‐derived markers for predicting cognitive and functional decline.

3.4.3. Chronic Kidney Disease

Shifting focus to another condition, chronic kidney disease (CKD) is highly prevalent, with low awareness, poor prognosis, and substantial medical costs. It is defined by impaired kidney function and a reduced glomerular filtration rate, a key indicator for staging CKD (Jha et al. 2013; Inker et al. 2014). CKD‐associated vascular injury and dysfunction may impair cerebral perfusion and disrupt iron homeostasis, potentially contributing to neurological complications (Girard et al. 2017).

Voxel‐wise QSM studies have shown stage‐dependent alterations: patients with stage 1‐3a and 3b‐5 exhibited higher magnetic susceptibility in the caudate nucleus and putamen, but lower values in the hippocampus compared to healthy controls (Figure 6c) (Song et al. 2024). In stage 5 CKD nondialysis patients, magnetic susceptibility was reduced in the hippocampus and thalamus compared to healthy controls, but elevated in the caudate nucleus relative to earlier stages (Wang et al. 2023). Notably, cerebral blood flow (CBF)‐magnetic susceptibility coupling is weakened in the hippocampus, thalamus, and caudate nucleus, suggesting impaired vascular–metabolic regulation. These findings suggest a complex interplay between kidney function, cerebral blood flow, and iron deposition in CKD.

Current research in CKD focuses on early biomarkers of cognitive impairment and clarifying the role of iron in mediating effects of uremic toxins such as indoxyl sulfate and trimethylamine N‐oxide (Capasso et al. 2025). Future studies should aim to clarify the mediating role of iron deposition in the association between these uremic toxins and cognitive decline in CKD patients.

3.4.4. Headache Disorders

Headache disorders encompass a spectrum of clinical entities, including conditions characterized by new daily persistent headache (NDPH) as well as migraine disorders. NDPH presents with persistent headache lasting more than 3 months, while migraine is characterized by recurrent, often unilateral or paroxysmal headache attacks and is further classified into episodic migraine (EM) and chronic migraine (CM) according to attack frequency and duration (Headache Classification Committee of the International Headache Society 2018; Lipton et al. 2015). Growing evidence indicates that abnormal iron accumulation may contribute to the pathophysiology of these headache disorders (Li, Zhao, et al. 2024; Fila et al. 2024).

Whole‐brain QSM analyses have revealed subtype‐specific alterations across headache disorders. EM patients show elevated magnetic susceptibility in the left putamen and bilateral SN compared to healthy controls, regions densely innervated by serotonin (Li, Zhao, et al. 2024). In contrast, CM patients are characterized by widespread cortical magnetic susceptibility increases, including the right precuneus, insula, supramarginal gyrus, dorsolateral superior frontal gyrus, postcentral gyrus, cuneus, and left postcentral gyrus (Chen, Dai, et al. 2021). Similarly, NDPH patients exhibit increased magnetic susceptibility in gyrus rectus, right insula, and right cerebellum lobule VIII (Chen, Pei, et al. 2025).

While other classifications (e.g., with or without aura) exist (Headache Classification Committee of the International Headache Society 2018), research on iron patterns across these groups remains scarce, underscoring gaps in understanding migraine pathogenesis.

4. Conclusion

QSM has emerged as a powerful neuroimaging technique, enabling the precise quantification of tissue magnetic properties across the brain. Whole‐brain voxel‐wise analysis allows an unbiased assessment of tissue alterations in subcortical structures, cortical areas, and white matter pathways without relying on a priori anatomical assumptions. This capability has significantly advanced our understanding of various neurodegenerative and neuropsychiatric disorders.

To support the reproducibility and translational potential of findings, this review highlights key methodological limitations, sources of variability across voxel‐wise QSM studies, and unresolved challenges related to normalization, statistical inference, and other methodological factors. Furthermore, the provided recommendations and reporting guidelines are intended to serve as a practical reference for future studies. As the field continues to evolve, these methodological considerations and advancements will further establish QSM as an indispensable tool for both clinical research and routine neurological assessment.

Funding

This work was supported by Shanghai Municipal Science and Technology Commission (Grant No. YDZX20243100003001001) and the Science and Technology Innovation 2030 Major Projects (Grant No. 2021ZD0200500).

Disclosure

No use of generative AI in scientific writing, analyzing, and drawing could have appeared upon submission of the paper.

Conflicts of Interest

Yi Wang owns equity of Medimagemetric LLC, a Cornell spinoff company. The other authors declare no conflicts of interest.

Supporting information

Data S1: hbm70471‐sup‐0001‐Supinfo.docx.

Tang, Y. , Zhong H., Li G., Wang Y., and Li J.. 2026. “Voxel‐Wise Whole‐Brain Analysis in Quantitative Susceptibility Mapping: A Narrative Review.” Human Brain Mapping 47, no. 3: e70471. 10.1002/hbm.70471.

Data Availability Statement

Data sharing not applicable to this article as no datasets were generated or analysed during the current study.