The development of diffusion magnetic resonance imaging (dMRI) coupled with the subsequent introduction of techniques to model water diffusion within the brain tissue using a diffusion tensor model (DTI) (Delmarcelle and Hesselink, 1992; Basser et al., 1994a, 1994b; Carter et al., 2015), has led to unprecedented opportunities for noninvasive exploration of the brain’s intricate white matter (WM) structures (Clayden et al., 2007). Tractography is a technique that harnesses data derived from dMRI to reconstruct and visualize the WM pathways within the brain by tracing the likely paths of water diffusion. Tractography (Mori et al., 1999; Basser et al., 2000) involves the algorithmic reconstruction of these WM pathways, generating a multitude of fibers (El Kouby et al., 2005) for each subject. This is followed by the delineation of the obtained fiber trajectories or streamlines into bundles or their association with anatomically well-defined tracts, a process commonly referred to as WM tract segmentation or dissection (Bullock et al., 2019).

WM tracts in the brain serve as the communication highways that connect different regions of the brain. Accurate segmentation enables researchers and clinicians to identify specific tracts associated with particular neurological functions, including cognitive, motor, and behavioral processes (Yushkevich et al., 2008; Sadeghi et al., 2013). Accurate tract segmentation plays a pivotal role in comprehending alterations in the micro- and macro-structure of the brain’s WM. It enhances our understanding of how structural connectivity shapes brain function and development. Additionally, it provides valuable insights into neurological diseases, including cognitive impairment and neurodegeneration, mental health disorders, and the aging process (Catani, 2006; De Belder et al., 2012; Le Bihan and Johansen-Berg, 2012). Moreover, accurate WM tract segmentation holds immense clinical significance, particularly in aiding in pre-operative and intra-operative brain tumor resections. It facilitates the visualization and localization of WM tracts that may be displaced or affected by tumors (Lazar et al., 2006; Chen et al., 2016; Essayed et al., 2017; Vanderweyen et al., 2020). It is worth noting that WM tract segmentation is a very challenging task. The human brain contains millions of intertwined axonal pathways, and these fibers can cross, split, or merge, making it challenging to accurately track individual pathways.

Most techniques employed for WM tract segmentation are based on virtual dissection or manual approaches, which involve the meticulous delineation of regions of interest (ROIs) (Catani et al., 2002; Mori and van Zijl, 2007; Wakana et al., 2007). These ROIs define where streamlines should pass and where streamlines should terminate. The provision of ROIs requires expert knowledge and hence manual methods incur expert labor costs. Manual methods face practical challenges in their adoption since they are time-consuming and expensive due to their high clinical and labor costs. Nevertheless, manual methods remain the gold standard for delineating WM tracts and serve as a critical benchmark for validating alternative approaches. The advent of better imaging techniques, improved image quality and higher resolutions (Van Essen et al., 2012), along with the application of sophisticated post-processing techniques, has driven a significant surge in the development of automated methods for tract segmentation (Yamada et al., 2009; Essayed et al., 2017; Ghazi et al., 2023).

A wide range of automated white matter tract segmentation methods have been developed over the years. While multiple works exist that review tractography methods and their applications, currently, there is limited literature available that specifically discusses the topic of delineating white matter tracts. Authors summarize the various categories that tractography segmentation methods fall under (Zhang et al., 2022) when reviewing quantitative tractography methods for studying the brain’s structural connectivity in health and disease. Recently, authors in Ghazi et al. (2023) have reviewed literature focusing on deep learning approaches for tract segmentation. In this work, we extend the scope by conducting a systematic and comprehensive review of automated approaches for the segmentation of white matter tracts in the last decade. This paper contributes to the following:

The remainder of the survey is organized as follows: Section 2 presents the review planning, Section 3 introduces the key findings as results, Section 4 summarizes and discusses the findings, Section 5 outlines the future directions, and Section 6 concludes the review.

This section is dedicated to planning the review: the comprehensive research questions related to the study are rigorously defined, the identification criteria and the resources of study are detailed.

Our search strategy retrieved 619 articles published between January 2013 and September 2023. After articles were reviewed for definite exclusions and the bibliography of eligible articles were hand-searched, 59 articles met the inclusion criteria. Figure 1 shows the flow diagram of the retrieved articles and the rules applied to get the resulting 59 articles. The results are presented as follows: First, we summarize the major datasets used in the studies included in this review. We then provide a list of the 59 research articles by focusing on the research questions established. These research articles are mentioned according to the categories they belong to and finally we provide a summary for the evaluation metrics used by the studies.

In this review paper, we have provided a systematic review of automated methods for white matter tract segmentation with respect to the most widely used public datasets for this task, the various categories of automated methods developed, and the evaluation metrics used to study the performance of the method. Although there are studies that have reviewed automated methods for brain tractography (Poulin et al., 2019; Zhang et al., 2022) and also deep learning methods for tract segmentation (Ghazi et al., 2023), to the best of our knowledge, a systematic review that focuses on automated methods for tract segmentation has not been published yet. This review paper underscores the methodological advancements in building automated methods, as evidenced through the 59 articles included in this review.

While manual segmentation of tracts or virtual dissection methods were not a focus of this survey, multiple approaches have been proposed in the last decade that conduct fiber selection and anatomical labeling using expert knowledge (Rheault et al., 2020, 2022a,b; Ille et al., 2021). These methods focus on improving the design of white matter dissection protocols to build more generalizable and reproducible methods. In Schilling et al. (2021) authors show the need to have a standard nomenclature and definitions for white matter bundles and that there are still issues in tractography segmentation that need to be resolved so that they can be used in routine clinical settings. Such methods are worth mentioning in this survey since they show that segmenting white matter tracts is a crucial task and that there is still a lot of scope for improvement.

It is also important to note that our study of white matter tract segmentation focusses on fibers in the deep white matter. Multiple studies are available that investigate the segmentation of subcortical U-fibers, which are special types of short association fibers located in the superficial white matter (Guevara et al., 2017, 2020; Xue et al., 2023). Despite studies on superficial white matter (SWM) being sparse due to its complexity (Xue et al., 2023) have employed point-cloud-based deep learning techniques that concentrate on superficial white matter tract segmentation. Also, numerous studies on automated white matter tract segmentation methods were omitted from this review because they did not meet the search criteria used to compile the literature included in this study. For example, studies (Bazin et al., 2011; Yendiki et al., 2011) provide automated methods for white matter tract segmentation, however, were not included in this study since they are published before 2013. Studies centered on automated fiber tracking or tractography (Teeuw et al., 2015; Warrington et al., 2020) are not featured in this review; however, they merit attention as they play a vital role in advancing new automated methodologies for accurately reconstructing white matter pathways, thereby facilitating the analysis of extensive datasets. Authors in Warrington et al. (2020) present tractography protocols as a software tool for standardized and automated cross-species tractography generated from large datasets. Automated tractography methods such as TRACULA (Yendiki et al., 2011), Teeuw et al. (2015) have used learning-based methods that show highly promising performance incorporating information on the anatomy of the pathways for reconstruction of white matter pathways thereby facilitating automated fiber tracking to large studies.

Twenty-seven percent of the papers in the current survey are based on direct voxel-based segmentation methods. Our results show that papers based on fully convolutional networks are typically based on encoder-decoder architecture such as U-net (Ronneberger et al., 2015). These voxel-based methods are gaining popularity with the advent of new and efficient deep learning-based segmentation techniques. However most current studies still rely on U-net based architectures as the baseline model, and the popular segmentation architectures like those based on transformers (Dosovitskiy et al., 2020; Hatamizadeh et al., 2021; Cao et al., 2022) have not been applied to this domain yet. Deep learning methods also perform segmentation by either labeling the streamlines or directly labeling the voxels. In general, the progress seen in using deep learning methods for medical image segmentation tasks (Hatamizadeh et al., 2021; Xiao et al., 2023) has not yet been fully applied to white matter tract segmentation. This is mostly because it is more demanding to have manual annotations of white matter tracts than other brain anatomical structures. Also, while deep learning methods provide fast segmentations, their results can still be unsatisfactory, and are not robust to changes of bundle sizes, tracking methods and data quality (Bertò et al., 2021). This shows that the major challenges in using machine learning or deep learning methods for tract segmentation will require researchers to come up with more generalizable solutions, create and publish more annotated datasets, use other techniques like transfer learning, self-supervised learning to overcome the challenges of limited training samples for deep learning-based methods to gain clinical applicability.

Next, our results show 25% of the papers in the current survey applied streamline-based clustering methods and 20% used streamline-based classification methods for automatic tract segmentation. These methods focus on clustering large number of fiber trajectories or streamlines into clusters or fiber bundles. However, few of these methods attach labels to clusters, and the clusters must be assigned labels either manually or automatically by using a streamline atlas usually incorporated in the clustering process. Such methods also have to post-process their results in order to filter the bundles to exclude spurious tracts that are falsely included in the clustering results.

In all the methods seen in this work, only two studies were found which used registration-based methods for white matter tract segmentation (Garyfallidis et al., 2018; Jordan et al., 2021). In Wasserthal et al. (2018) authors compared their work with two registration-based methods for automatic tract segmentation, which usually involves using a tract atlas and registering it to the subject of interest which yields a binary mask for each tract in subject space. Fewer registration methods are likely employed due to the inaccuracies produced during the registration step, and the computational complexity needed. However recently, a lot of work has been done in using deep learning methods for image registration (Oliveira and Tavares, 2014; Fu et al., 2020) to overcome the challenges of traditional registration methods. For example, authors in Zhang et al. (2021) proposed a deep learning-based method for registration of dMRI images. This exemplifies the growing interest in applying such methods to tract segmentation. Currently, to the best of our knowledge, there are no studies that use deep learning-based registration techniques for the task of automated white matter tract segmentation.

It is interesting to note that there is limited existing work on automated methods for segmenting white matter bundles for the neonatal brain. In our survey we only found two papers (Ratnarajah and Qiu, 2014; Logiraj et al., 2021b) that correspond to this topic. This could mainly be because segmenting white matter structures is particularly difficult in the neonatal brain since it is undergoing a critical growing process along with cellular maturation such as myelination and synaptic pruning (Ratnarajah and Qiu, 2014). Existing methods rely mostly on fully manual segmentation for delineating white matter structures (Oishi et al., 2011) or are based on semi-automated techniques (Huang et al., 2006). In Oishi et al. (2011) authors developed an atlas-based segmentation based on image registration, which also needs manual expert assessment in order to delineate the required white matter structures. This work was developed almost a decade ago and there have been multiple automated segmentation techniques proposed since that have been successfully applied to adult’s brain as shown in this survey. Manual methods also suffer challenges of being time consuming and require prior anatomical knowledge to achieve reasonable accuracy and reproducibility.

Overall, we observe that automated tract segmentation algorithms follow varied methods for pre-processing, augmenting, and training their datasets and few methods use multi-site datasets. Even the techniques used to generate reference tracts are not the same across most of the methods. This makes it impossible to assess the true generalizability and reliability of the proposed methods (Poulin et al., 2022). This problem is also observed in manual segmentation methods where there is varied reproducibility for segmenting the same tracts among different experts or the inter-protocol agreement across protocols for various white matter pathways is poor as shown in (Schilling et al., 2021).

We observe that most studies included in this review do not provide computational time making it challenging to assess the practicality of these methods. In general, streamline-based methods typically require substantial memory for generating millions of streamlines per subject, whereas direct voxel-based methods can segment white matter tracts for a test subject in under a minute (Wasserthal et al., 2018). Out of 59 segmentation methods reviewed, only 18 have been validated on external datasets with varying scanners and acquisition parameters. This lack of generalizability testing may be due to the limited availability of publicly accessible tract segmentation datasets. Despite this, direct voxel-based methods can be used for data augmentation during training to simulate domain shifts in external datasets, potentially reducing the domain-shift impact.

Lastly, we have summarized the most common evaluation metrics used by tract segmentation methods to validate their results in Section 3.3. However, there is no consensus on the evaluation metrics used to compare the various proposed approaches. Due to the limitation of ground truth, most methods rely on reproducibility in terms of intra- and inter-rater as well as test–retest reproducibility (Zhang et al., 2019; Rheault et al., 2020, 2022b) and consistency of methods across different populations and acquisitions (Wasserthal et al., 2018, 2019), as validations points for identifying a good tract segmentation method.

Although this paper reveals the advancements of automated methods for white matter tract segmentation, there is still a lack of a general standardized method that can be reliably used by clinicians. There is still limited consensus on the definition of tracts even among knowledgeable and experienced professionals who are concerned about the inter- and intra-user reproducibility with manual placement of ROIs (Zhang et al., 2010). This further complicates the methodology development and validation process. This suggests that there is a need for the development of more standardized approaches for validation tract segmentation results.

The recent work of authors of TractSeg (Wasserthal et al., 2018) enabled the distribution of manually labelled tracts to the community so that researchers could collaboratively share the segmented tracts by experts. This gave rise to the development of more generalized approaches towards white matter tract segmentation, which otherwise would not have been possible. This has set a particularly good example so that in the future, researchers can continue to enable the progress, development, and assessment of higher-quality automated methods through such public collaborations. There is still substantial room for future improvements in the domain of generating high quality ground truth via expert neuroanatomists.

Another important aspect to consider when developing automated methods is their computational cost. With the advent of improved imaging tools for the acquisition of data and increasing efficiency of computational resources, there is a critical need for building applications that can be clinically used. Moreover, there has been a significant surge in image sizes. A decade ago, state-of-the-art MR acquisitions typically featured MR images of human brains with voxel sizes of 2 × 2 × 2 mm³. Today, we routinely encounter voxel sizes smaller than 1 × 1 × 1 mm³, as seen in data collected by projects like the Human Connectome Project (Van Essen et al., 2012; Glasser et al., 2016). Therefore, rapid tract segmentation approaches are needed to allow interactive analysis and also to efficiently handle very large imaging studies in a time and cost-effective manner.

Another important future direction would be to consider tractogram data generated from varied tracking algorithms as input to the automated tract segmentation methods developed. This is because a variety of tracking algorithms with different parameter values can be used by tractography studies to generate the tractograms. Then the segmentation of tracts could be applied to any of these generated tractograms, and the method should be able to adapt to all these diverse types of inputs.

This systematic review summarized 59 relevant articles in all. Unlike previous studies, our work focuses on a systematic review of methods for automated white matter tract segmentation developed in the last decade. This work framed crucial research questions to explain what approaches have been used for automated tract segmentation methods, discover key research gaps, determine datasets that are publicly accessible for researchers and summarize the most common evaluation techniques utilized. The literature published in this area as displayed and characterized in the Results section is one that is of growing and global interest.

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

AJ: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. HL: Investigation, Project administration, Resources, Supervision, Writing – review & editing. NP: Funding acquisition, Project administration, Resources, Supervision, Writing – review & editing. LH: Funding acquisition, Project administration, Resources, Supervision, Writing – review & editing.