Introduction

The human brain is the most complex organ in terms of structure and function, making brain atlases essential tools for studying its intricacies.[1–6] Brain atlases not only enable researchers to understand complex functional circuits and the neural basis of cognitive behaviors but also provide insights into development mechanisms, identification of early diagnostic biomarkers, and the establishment of personalized and precise treatments for brain diseases.[7,8] In the past decade, neuromodulation techniques[9,10] such as transcranial magnetic stimulation (TMS)[11] and deep brain stimulation (DBS)[12] have rapidly advanced, enabling personalized treatments for various neuropsychiatric disorders. These techniques can improve patient outcomes by precisely targeting specific brain regions or circuits, altering neuroelectrical activity and transmitter release.[13,14]

Recent advancements in neuroimaging and computer-assisted intervention have led to new concepts like stereotactic neurosurgery and neurosurgical navigation systems. While these technologies have made substantial advancements in the treatment of brain diseases, they are confronted with a formidable challenge—the pronounced variability in pathological mechanisms and individual patient differences shaped by a complex interplay of genetic and environmental factors.[15] This interplay results in structural disparities within the brain, encompassing differences in size, shape, cellular architecture, and brain connectivity. The uniqueness of each individual’s brain structure extends its influence to brain function, leading to variability at behavioral levels, including functional signals,[16] cognition,[17] psychiatric symptom,[18] and clinical neurosurgery.[19,20] As a result, there is a need for personalized navigation solutions and targeted neuromodulation strategies for each disease and patient. Current commercial brain navigation systems often rely on traditional anatomical brain atlases, which have limitations due to their basis on postmortem specimens, insufficient information on individual variability, and lack of functional brain parcellation.[7,21] This hinders clinical effectiveness and the precision of neuromodulation therapies. A major bottleneck in navigation-based neuromodulation lies in the lack of fine-scale brain atlases and individualized approaches, limiting the development of personalized and precise treatments. Existing stereotactic navigation systems often require manual intervention, reducing efficiency and introducing potential subjective errors. Consequently, it is crucial to integrate and analyze individualized brain mapping methods for specific neuromodulation navigation, thus to better characterize pathological changes, understand neuromodulation mechanisms, and provide a basis for early diagnosis and prognostic evaluation.

In this review, we presented an overview of recent advances linking individualized brain mapping and navigated neuromodulation, focusing on establishing precise neuromodulation paradigms based on individualized brain mapping. We discussed the latest approaches to individualized brain mapping, their methodological profiles, advantages, disadvantages, and application trends. We also reviewed the progress in non-invasive neuromodulation technology, providing an overview of current TMS targeting methods and future directions. Furthermore, we illustrated how various DBS procedures localize stimulation targets and how the latest individualized brain mapping techniques promote DBS localization. Finally, we summarized our findings and posed open questions about the future directions of individualized brain mapping navigated neuromodulation.

Current Status of the Methodology for Individualized Human Brain Mapping

From initial attempts a hundred years ago[1] to recent advancements,[7,8] individualized brain mapping has experienced several stages as brain imaging techniques have evolved. The pioneering method[22] involved constructing a brain atlas using postmortem brain tissues from a single subject and manually labeling brain regions based on cytoarchitectural features. The most widely utilized atlas among these is the Brodmann atlas.[1,22] This first specimen atlas highlighted the potential of individualized brain mapping, but brain atlases based on an individual’s anatomical characteristics lacked population commonality and provided only a coarse division of brain regions. More recent studies have employed staining techniques to map specimen brains at the group level, using cytoarchitectonic and myeloarchitectonic characteristics[4,5] extracted from multiple subjects. The increased number of subjects improved the population coverage of the brain atlas; however, anatomical annotation is labor-intensive, and the experience and annotation standards of annotators are not uniform, making it challenging to create a group-consistent brain atlas. Furthermore, since ex vivo manual annotation is invasive, histological brain atlases cannot be reproduced in living subjects. In this case, various individualization techniques [Figure 1, Table 1] have been employed in individualized brain mapping.

Figure 1: Schematic overview of individualized brain mapping techniques. A detailed description can be seen in Table 1. Table 1 - Summary of individualized brain mapping techniques. Categories Items Descriptions Schemas Registration-based individualized brain mapping Structure registration[23] Minimizing local structural differences between reference and individual atlases Figure 1: structure registration Diffusion registration[24] Minimizing differences in local diffusion characteristics or structural connectivity between reference and individual atlases Figure 1: diffusion registration Function registration[25] Minimizing differences in functional activation or functional connectivity patterns between reference and individual atlases Figure 1: function registration Multi-modality registration[26] Minimizing feature differences between reference and individual atlases across multi-modalities Figure 1: multi-modality registration Multi-atlas registration[28] Minimizing feature differences between multiple reference atlases and individual atlas Figure 1: multi-atlas registration Unsupervised learning-based individualized brain mapping Boundary mapping[6] Not limited to data modalities, but often used in fMRI to find where changes are sharpest as brain region boundaries Figure 1: boundary mapping Region growing[30] Not limited to data modalities, brain regions start at random centroids and expand outward until convergence to boundaries Figure 1: region growing Individual clustering[31–33] Not limited to data modalities, the mainstream unsupervised clustering methods are K-means, hierarchical clustering, and spectral clustering Figure 1: individual clustering Community detection[34] Not limited to data modalities, a graph theoretic approach is utilized to perform subgraph cuts to delineate brain regions Figure 1: community detection Group prior-guided individualized brain mapping Tractography projection[35] Only used for dMRI data, relying on reference atlas to offer seed brain regions for tractography between cortical and subcortical areas Figure 1: tractography projection Decomposition[36,37] Mostly used for fMRI data, utilizing a group of subjects to build the group-level reference components and then projecting it onto individual subjects Figure 1: decomposition Exemplar-based clustering[38,39] Mostly used for fMRI data, utilizing a group of subjects to build the group-level exemplar map and then performing affinity propagation clustering for individuals Figure 1: exemplar-based clustering Boundary iterative adjustment[40,41] Not limited to data modalities, utilizing a group of subjects to build a reference probability atlas and then iteratively adjusting the region boundary in individuals until convergence Figure 1: boundary iterative adjustment Probabilistic modeling[42,43] Mostly used for fMRI data, utilizing a group of subjects to optimize inter-subject, intra-subject, and inter-region variability and to build the individual atlas Figure 1: probabilistic modeling Deep learning[44,45] Not limited to data modalities, utilizing a group of subjects to train an individualized brain mapping model and then predicting the parcellation pattern for individuals Figure 1: deep learning

dMRI: Diffusion magnetic resonance imaging; fMRI: Functional magnetic resonance imaging.

Registration-based individualized brain mapping

The most straightforward approach for individualized brain mapping is image registration. Regardless of the magnetic resonance imaging (MRI) data modality, the goal of single-modality registration is to align the reference atlas[21] to individual space and minimize reference-individual differences. The structure registration-based individualized brain mapping [Figure 1, Table 1] can depict local anatomical architecture like brain shape and volume size, but hardly assesses diffusion characteristics or functional activation patterns.[23] Diffusion registration-based individualized brain mapping [Figure 1, Table 1] can preserve diffusion characteristics, such as diffusion orientation and density, and is advantageous for tractography.[24] Function registration-based individualized brain mapping [Figure 1, Table 1] is to align functional MRI signals, ensuring a group-consistent functional activation pattern between subjects.[25]

The single-modality MRI registration-based individualized brain mapping depicts the parcellation pattern from a structural or functional perspective. To achieve a more comprehensive individualized brain mapping, multi-modality registration techniques were developed to integrate anatomical architecture, connectivity, and functional information of the cerebral cortex [Figure 1, Table 1] multi-modality registration).[26] In addition to the similarity constraint within the anatomical structure, multi-modality registration aims to minimize structural connectivity or functional connectivity differences between the reference and individualized atlases, allowing multi-modality features to be fused during the registration process. Compared to classical volume-based single-modality MRI registration techniques, surface-based multi-modality registration better preserves the biological features of the cortex.[26,27] However, surface-based multimodal registration is limited by substantial surface shape variability between individuals.[26] The structural connectivity and functional networks often differ even in the same cortical placement between subjects.

In addition to multi-modality registration techniques, multi-atlas registration [Figure 1, Table 1] offers another approach to individualized brain mapping.[28] While multi-modality registration combines parcellation characteristics from different MRI modalities, multi-atlas registration aims to integrate parcellation patterns from a group of reference brain atlases. Briefly, multi-atlas registration assigns reference atlas labels with the highest likelihood to a new subject, thereby constructing the individualized brain atlas. A typical multi-atlas registration framework includes reference atlas generation, atlas registration, label propagation, and label fusion.[28] Notably, the multi-atlas registration-based individualized brain mapping is sensitive to label fusion algorithms. Individualized brain mapping based on multi-atlas registration is also capable of integrating parcellation patterns from cross-modality reference atlases.[29]

In summary, while single-modality MRI registration methods could construct a non-invasive individualized brain atlas compared to histological labeling, they have insufficient parcellation perspective. Multi-modality MRI registration methods offer a comprehensive perspective by integrating cytoarchitecture, connectivity, and functional aspects into individualized brain mapping, but they encounter individual variability issues. Multi-atlas registration techniques provide a more flexible way to individualized brain mapping by fusing parcellation patterns from a group of reference atlases with single or multiple modalities. In the future, registration-based individualized brain mapping is likely to lean toward cross-modality multi-atlas registration, integrating with the burgeoning deep learning technology to develop a range of clinical applications.

Unsupervised learning-based individualized brain mapping

While image registration-based individualized brain mapping techniques have been widely used in clinical applications,[11,12] they have limited ability to accurately capture individual specificity. To better capture the individual specificity in brain mapping, unsupervised learning-based methods rely solely on the subject’s own MRI data. The ultimate purpose of unsupervised learning-based individualized brain mapping is to segment structurally or functionally integrated brain tissues into segregated brain regions.[7] To achieve this goal, several unsupervised learning-based methods have been developed, which can be classified into four categories: boundary mapping,[6] region growing,[30] individual clustering,[31–33] and community detection.[34]

Inspired by edge detection algorithms in image segmentation studies, boundary mapping was initially used to identify locations where parcellation features change rapidly in the target ROI [Figure 1, Table 1].[6] These abrupt local changes in cytoarchitecture, connectivity, or function form the biological basis of boundary mapping.[6] Consequently, the boundary mapping method does not require setting the number of subregions, which is automatically determined by the distribution of the parcellation features. However, constructed brain atlases may be inconsistent when using different parcellation features due to variations in the distribution of boundary features. Notably, most unsupervised learning-based individualized brain mapping methods are not limited to parcellation features and MRI modalities.[7] Compared to seeking local boundaries where parcellation features change rapidly, the region growing method[30] [Figure 1, Table 1] starts from the central seed points in the target ROI and iterates outwards. The iteration is to find the voxels with the parcellation features most similar to the seed points or regions and merge them into the updated seed regions until the iteration traverses all voxels. These iterations are multi-channel, iterating from multiple seed points simultaneously. Thus, the parcellation granularity of the region growing method is determined by the number of seed points. Additionally, the final individualized brain atlas may be affected by the location of the initial seed points. Methodologically, boundary mapping and region growing are complementary, with the initialization of the former being the boundary and the latter being the central points.

Typical individual clustering algorithms applied in individualized brain mapping are K-means, hierarchical, and spectral clustering [Figure 1, Table 1]. K-means clustering assigns voxels to several centroids according to the principle of nearest distance, thereby assigning all voxels of the target ROI to a given number of clusters.[31] Hierarchical clustering aims to build a hierarchical relationship map according to the similarity between voxel or cluster pairs.[32] Spectral clustering first reduces the dimension of the parcellation features and then performs K-means clustering on the reduced-dimensionality feature matrix.[33] Community detection is a type of graph segmentation algorithm regarding the target ROI as an adjacency graph [Figure 1, Table 1].[34] The purpose of a community detection algorithm is to segment an adjacency graph into subgraphs by minimizing the distance within the subgraph and maximizing the distance between the subgraphs. Clustering pays more attention to the inherent attributes of nodes, whereas community detection focuses on the connections between nodes. Also, while clustering methods are sensitive to cluster numbers, community detection algorithms can estimate the optimal number of subgraphs.

Boundary mapping or region growing methods focus on local variability or continuity, while individual clustering and community detection methods pay more attention to the global relationship. Utilizing the subject’s own MRI data rather than reference atlases, these unsupervised learning-based individualized brain mapping methods are advantageous in depicting individual specificity. In the future, there is an urgent need for unsupervised learning-based individualized mapping methods that can work with low-resolution diffusion MRI (dMRI) and short-scan functional magnetic resonance imaging (fMRI), which are commonly acquired in clinical settings. Additionally, as high-resolution dMRI and long-scan fMRI become more accessible, unsupervised learning-based individualized brain mapping will empower the neuroscience community to achieve substantial advancements beyond the current state.

Group prior-guided individualized brain mapping

Unsupervised learning-based individualized brain mapping methods effectively capture individual specificity in parcellation patterns, but their moderate population commonality, due to the lack of prior information from reference atlases, leads to challenges in depicting inter-subject consistency. Moreover, unsupervised learning-based methods,[7] with different settings, may produce highly variable parcellation results because of the absence of constraints from prior information. To address these issues, group prior-guided individualized brain mapping methods were proposed to integrate reference atlases with individual MRI data. Current group prior-guided individualized brain mapping methods include tractography projection,[20,35] decomposition,[36,37] exemplar-based clustering,[38,39] boundary iterative adjusting,[40,41] probabilistic modeling,[42,43] and deep learning.[44,45] These approaches employ prior information from group-level reference brain atlases to constrain or initialize the construction of individualized brain atlases. Consequently, group prior-guided individualized brain mapping can capture both high individual specificity and population commonality.

The tractography projection method depends on registering the reference cortical atlas to the individual brain and performing tractography between the subcortical seed ROI and cortical reference brain regions [Figure 1 & Table 1].[35] The ROI is then divided into reference-related functional zones based on the fiber connectivity profile. As fiber bundles connect functional circuits between cortical and subcortical regions, tractography projection-based individualized brain mapping aligns with neurological interpretation. This approach has been applied in localizing DBS targets.[20] However, the brain atlas constructed by tractography projection relies on an accurate reference cortical atlas and robust tractography algorithms. The reference cortical atlas is of great importance in giving credible interpretations of the delineation results. Since the reference cortical atlas is constructed by data from a group of subjects, it has strong population commonality but weak individual specificity. Whereas subjects differ because of their pathological states and brain anatomy, using the reference atlas to substitute for the individual’s brain parcellation pattern may lead to inaccurate brain mapping results. Hence, developing an accurate cortical brain atlas of an individual subject is a necessary step before building an individualized subcortical atlas by tractography projection. In terms of tractography algorithms, false positive fiber bundles and fiber crossing issues may also compromise the tractography-based projection’s performance. To enhance the tractography algorithms, high spatial resolution dMRI data has been utilized to create tractography projection-based individualized brain mapping.[20] In the future, the availability of clinically oriented high-field MRI technology may further advance this individualization method.

The decomposition methods first generate group-level components [Figure 1, Table 1],[36,37] then calculate each subject’s loading matrix on the group-level components, and finally map the loading matrix to the corresponding brain spatial location to create an individualized brain atlas. But there are negative loadings that may not be biologically meaningful. To prevent negative loadings, a group-guided non-negative matrix factorization (NMF) method was proposed.[37] However, the individual loading matrix’s sensitivity to group-level component generation makes group prior-guided decomposition methods vulnerable to the number of subjects and the dimension of parcellation features.[37] More recently, an autoencoder model has been employed to extract group-level components in a non-linear manner, providing an individualized brain atlas with flexible parcellation granularities.[45]

Exemplar-based clustering builds upon traditional individual clustering while incorporating prior information from reference atlases [Figure 1, Table 1]. Researchers have used unsupervised learning-based individual clustering methods to construct an individualized brain atlas based on data from a specific subject; however, this approach may produce an individualized brain atlas that is sensitive to initial centroids and cluster numbers. To address this issue, exemplar-based clustering first identifies the most representative exemplars of a group of subjects using a greedy algorithm.[38] Then, it assigns a voxel to the exemplar-corresponding clusters by minimizing the distance to the determined exemplars. Similarly, group-guided affinity propagation clustering first finds the cluster centers and the optimal number of clusters in a group-level connectivity matrix.[39] Then, it uses the group-level results to initialize individual-level affinity propagation clustering. While exemplar-based clustering individualized brain mapping methods are robust in constructing highly individual-specific and inter-subject consistent brain parcellations, they are time-consuming and computationally complex.[38,39] Consequently, these methods may be more suitable for individualized brain mapping with low-dimensional parcellation features or small-volume ROIs. In the future, dimensionality reduction algorithms and parallel computing techniques may be employed to reduce computational burden and accelerate convergence.

Iterative boundary adjustment first creates a group-level reference atlas based on a large population dataset [Figure 1, Table 1].[40] The group-level reference atlas is then used as an initialization of individualized brain mapping. Each vertex or voxel is iteratively reassigned to the functional network with the most similar parcellation features until the iterative adjustment process converges according to a predefined criterion. This method is robust whether the parcellation features are functional connectivity[40] or structural connectivity.[41] Since the parcellation pattern of an iterative adjustment-based individualized brain mapping depends on a reference atlas, it is essential to construct canonical reference brain atlases.[21] Currently, the main technical challenge of iterative boundary adjustment methods is portability, as many empirical experiments are required to find the optimal termination criteria in different application scenarios. Furthermore, reference atlases constructed by normal subjects may not accurately reflect brain atrophy or dysconnectivity in patients. In the future, group-level age-specific and disease-specific brain atlases may enhance the clinical applicability of the boundary iterative adjusting method.

Probabilistic modeling methods use group-level parcellation features and labels as a prior probability distribution [Figure 1, Table 1].[42,43] To construct an individualized brain atlas for a new subject, the maximum posterior probability is calculated by maximizing the consistency within subregions and the difference between subregions. Earlier probabilistic modeling-based atlas focused on population-level parcellation patterns and inter-region variability. Later, inter-subject variability was introduced to the calculation of posterior probability.[42] Coupled with advancements in high-quality fMRI data, multi-session scanning has enabled researchers to consider intra-subject variability.[43] Probabilistic modeling methods are primarily used for fMRI data and allow the fusion of local and global parcellation features. Currently, the latest multi-session hierarchical Bayesian model (MS-HBM) has achieved impressive results for studying individual specificity at lab level.[43] However, applying the MS-HBM in clinical settings is challenging due to the requirement for high-quality MRI data, which necessitates long scanning times. In the future, data generation techniques may offer clinical feasibility for high-quality fMRI-based probabilistic modeling individualization methods.

Deep learning methods train classification or regression models using reference brain atlases and signals, and then test the trained models using a new individual’s data to construct an individualized brain atlas (Figure 1 & Table 1).[44,45] These deep learning-based individualized brain mapping methods are not sensitive to parcellation features and data modality. In addition to benefiting from the powerful representation capacity of deep learning models, deep learning-based individualized brain mapping has been reported to be more capable of capturing individual variability than parameter-based group prior-guided methods.[44,45] However, as a “black box”, the parcellation criteria of deep learning models are only weakly interpretable from a biological perspective. Moreover, the accuracy of deep learning models is data-hungry and relies heavily on the number of training samples and model complexity. Excessive attempts at adjusting hyper-parameters and overly complicated models may lead to overfitting. In the future, deep learning-driven individualization methods may be used in neuroscience research on large-scale datasets. Simultaneously, with the development of super-resolution technology, individual-level deep learning models, such as self-supervised learning, may become dominant in the field of individualized brain mapping.

By integrating individual parcellation features and reference atlases, group prior-guided individualized brain mapping methods can effectively capture both individual specificity and population commonality. These methods have not only made significant contributions to neuroscience research but have also demonstrated clinical feasibility in some cases. In the future, advancements in dimensionality reduction, parallel computing, population-specific brain atlases, data generation techniques, super-resolution MRI, and self-supervised learning will further accelerate the development of clinically useful group prior-guided individualized brain mapping.

Summary

In the present study, individualized brain mapping methods were divided into three categories. Registration-based individualization methods preserve population commonality to the greatest extent, while unsupervised learning-based individualization methods capture individual specificity to the greatest extent. By combining group priors with individual characteristics, group prior-guided individualized brain mapping methods retain not only high population commonality but also capture sufficient individual specificity. These three types of individualization methods are well-developed and have been applied to various neuroscience research and clinical scenarios. In the future, several techniques may facilitate the development of individualized brain mapping. First, MRI data quality enhancement techniques will be crucial, as MRI data quality directly affects the performance of individualization methods and the accuracy of the constructed atlases. However, clinical MRI data often have poor spatial resolution and signal-to-noise ratios. Enhancing clinical data quality through algorithms can promote the efficiency of individual parcellation feature extraction. Second, the development of population-specific group atlases is essential. Most current reference atlases provide population-level features for healthy young individuals but largely ignore age variability and pathological differences in broader human brain populations. Establishing group-level brain parcellation patterns for different age groups and pathological states is vital, as aging and pathological conditions often result in nuclear atrophy and decreased fiber bundle connections, leading to unstable clinical localization data. Finally, individual-level individualization modeling will be a future direction for individualized brain mapping. Most current individualization models are at the group level, with all subjects sharing the same set of parameters in individualized brain mapping. This approach inevitably leads to spatial registration errors between the group prior and individual features, regardless of the registration method applied. Moreover, group optima in the pipeline parameters do not reflect individual-level optimal parameters for parcellation. A “one subject, one model” approach, providing subject-specific parcellation models as well as subject-specific parcellation pipeline parameters for each subject, will be the future direction of individualized brain mapping. In summary, individualized brain mapping is expanding and will continue to integrate into clinical practices, with even more exciting advancements expected in the near future.

Non-Invasive Neuromodulation Method: Electric Field Simulation Guided and Brain Circuits Targeted TMS

The previous section has focused on the state-of-art methodological review of individualized brain mapping [Figure 1, Table 1]. In the next section, we will discuss how individualized brain mapping approaches can be applied to the field of neuromodulation, specifically with regard to non-invasive neuromodulation using TMS. TMS is an established, safe, and effective non-invasive neuromodulation method that delivers focused magnetic pulses to the scalp, generating an electric field in the cortex and modulating neuroactivity in the targeted brain region.[46] TMS has been used to treat many neuropathic conditions, including depression,[47] migraine,[48] and obsessive-compulsive disorder[49] according to the US Food and Drug Administration (FDA). A specific TMS target that can significantly affect the brain circuits is required to achieve substantial therapeutic effects. This section focuses on individualized TMS targeting methods with depression being the primary focus because it is one of the most extensively studied diseases. Although the traditional therapeutic approaches, such as medication and psychotherapy, are effective in treating depression, they do not work for approximately 20%–30% of patients diagnosed with major depression disorder.[50] TMS was therefore proposed as an alternative therapeutic approach that directly affects neural activity and modulates neuroactivity in ways that differ from conventional approaches.[47