Tracking the visual system—from the optic chiasm to primary visual cortex

The optimal MRI dataset acquired for the purpose of tracking of the visual system should comprise (a) T1-weighted (Fig. 3a), and optionally a proton density (PD) map, to allow for the comprehensive delineation of brain tissue and structures, and (b) DW images necessary for the tractography (Fig. 3b). In the dMRI analysis, the T1-weighted images and PD maps are used in order to identify and segment regions of interest (ROIs) and white matter mask, which enhances tractography. The DW images provide information about microstructural properties (Fig. 3c) and what is critical for tractography, spatial organization of neural fibers (additionally, tractography may also provide its own contrast as shown in Figure 3d). The basic underlying concept is the measurement of dephasing of excited protons along the given direction simultaneously in all of the brain voxels [37]. The contrast in a DW image depends on the properties of diffusion processes occurring in the tissue and is controlled by two parameters set for MR scanning: b‑value (determines the sensitivity to diffusion; b0 is the lowest, where b‑value equal to, e.g., 3000 s/mm² is considered high) and b‑vector (describes the direction along which the diffusion is measured). The selection of DW acquisition parameters is a complex issue, as one needs to consider MRI scanner capabilities, the focus of study (imaging of the whole brain or a single structure), data quality requirements (signal-to-noise ratio, spatial resolution, number of b‑values, and gradient directions) and external factors, e.g., scanning time (detailed in: [22, 40]). An efficient alternative solution is the adaptation of already validated and well-established protocols such as the Human Connectome Project (HCP) acquisition protocols [42]. Publicly available datasets should also be considered an option to tune the analysis procedures (see Table 1).

Apart from the data acquisition parameters, it is important to consider the choice of tools for data preprocessing and tractography (see Table 2). Diffusion-weighted data preprocessing is an important stage, as it determines the data quality and may affect the final outcome of the study. While the applicability and/or necessity of constituent preprocessing steps depends on parameters of the data and specifics of the study, the preprocessing pipelines applied in the HCP project [16] are a commonly recognized standard. Table 3 provides a list of recommended preprocessing steps, created by extending the HCP preprocessing pipeline by more recent utilities, which result in a corrected DW dataset aligned to the T1-weighted image (Fig. 4). A recently emerging alternative to data preprocessing and analysis on local machines are online platforms, such as BrainLife [1]. BrainLife allows one to upload one’s own data, or to use existing publicly available datasets (which are either stored directly on BrainLife or can be downloaded from other repositories) and perform neuroimaging analysis in the cloud using online services.

If required, the currently recommended strategy would be a manual delineation based on T1-weighted images. This may change with emerging DL-based methods, which demonstrate good performance [25], but are not yet widely available.

A FreeSurfer software, regarded as a standard tool in automated brain segmentation, is capable of segmenting the optic chiasm, yet its results can be suboptimal (Fig. 5b, top). As such it is recommended to segment the optic chiasm manually in order to achieve optimal accuracy (Fig. 5b, bottom)—an approach widely used in optic chiasm imaging [30]. Similar to optic nerves, current DL developments allow for the segmentation of the optic chiasm, but they are still not widely tested and available—especially in the case of optic chiasm malformations.

The segmentation of the LGN is particularly challenging, as this structure does not stand out in T1-weighted images. While it is possible to use automated segmentation tools (e.g., FreeSurfer, SPM, FSL, GIF, MALP-EM, Mipav) to this end, this approach has limitations, e.g., LGN segmentation can even vary between versions (as shown for FreeSurfer in Fig. 5c). Furthermore, the lack of validation undermines the study’s reliability. Therefore, it is recommended to obtain PD maps during the data acquisition, as their contrast allows for an LGN identification. In the absence of this information, an alternative strategy is the tractography-based identification of the LGN [28]. We propose that this strategy would be even further improved by simultaneous tractography from both the optic chiasm and V1 with the target in the thalamus. The intersection of the streamline endpoints might serve to identify the LGN, which can be combined with the previously introduced FreeSurfer segmentation.

For the identification of the primary visual cortex (V1), three approaches are available: (1) the use of automated segmentation tools, (2) individualized retinotopic mapping, and (3) the estimation of retinotopic maps from anatomical priors. (1) Automated segmentation tools in principle allow for an estimation of V1; however, individual differences from the general template introduce errors in the results of the V1 ROI definition. Further, this approach does not provide an estimate of the retinotopy of V1. (2) The alternative is to obtain retinotopic maps of the visual cortex specifically for the respective individuals with dedicated functional magnetic resonance imaging (fMRI) measurements [32]. Due to the retinotopic organization of the visual cortex, visual areas can be identified via fMRI-based retinotopic mapping (reviewed in [18, 19, 43]). T2*-weighted BOLD gradient-EPI sequences are acquired in fMRI scans during visual stimulation, typically with a contrast-inverting or moving checkerboard pattern section that travels through the visual field in a systematic manner to create a specific spatiotemporal response pattern on the visual cortex. Conventional phase-encoded retinotopic mapping [33] or the neuro-computationally more demanding population receptive field mapping [13] can be applied to obtain cortical maps of the eccentricity and polar angle representations of the visual field. These are typically visualized on the computationally inflated or flattened surface of the visual cortex as derived from high-resolution T1 images. Visual area boundaries can be delineated from these maps [33] as depicted in Fig. 5d (top row) for the primary visual cortex (V1). (3) Based on the evidence of qualitatively consistent organization of primary (V1) and extra-striate (V2 and V3) visual area topography [12], Benson et al. [4] demonstrated the ability of an anatomical template to predict the retinotopic organization of the visual cortex with high accuracy using only a participant’s brain anatomy, i.e., the patterns of the gyral and sulcal curvatures. In the absence of retinotopic mapping data, a viable alternative is therefore the use of Benson’s atlas, which utilizes known anatomical priors in order to estimate V1 location, as well as polar angle and eccentricity map (see Fig. 5d, middle row). This information can be used to specify ROIs for the purpose of tractography.

In order to enhance the accuracy of the tractography it is recommended to use the known anatomical priors, such as the limitation of the reconstructed pathways to white matter only. Such white matter masks can be extracted using, e.g., FSL or FreeSurfer software. This approach can be further extended by taking into account other types of tissues, such as in five-tissue-type segmentation (implemented in MRtrix), which is a critical component in their proposed anatomically constrained tractography (ACT; [34]).

Basic, yet successful and well-established, is the DT [3] model (Fig. 6b, right), which’s popularity caused confusion of Diffusion Tensor Imaging (DTI; application of DT model) with the general dMRI term (covering all models). The DT model represents diffusion as a 3 × 3 matrix with six independent terms (i.e., from at least six volumes with unique gradient directions), which can be graphically represented as an ellipsoid. This model has limitations, as it models only one diffusion direction per voxel, and as such fails to represent more complex structures, such as the optic chiasm (due to the presence of crossing fibers), Meyer’s loop (due to its curvature), or posterior parts of optic radiation (due to fanning).

Among other alternatives (such as Q‑ball imaging, diffusion spectrum imaging), we would like to discuss in detail CSD [39], which treats signal in each voxel as a convolution of the response from a single fiber population and distribution of the local fiber’s orientations. As such, CSD is capable of resolving multiple fiber bundles crossing a single voxel, which are described by orientation distribution functions (ODFs). The fitting of the CSD model requires more than six gradient directions (typically 30–60), which grants noise reduction and higher angular resolution at the cost of longer scanning time.

Generally, there are two classes of tracking algorithms—deterministic and probabilistic. Deterministic algorithms assume that fibers in each voxel are oriented in only one direction as determined by the given model (see Fig. 6c)—as such they offer robust results, but fail to grasp complex architectures. Alternative probabilistic algorithms at each step of tracking sample the final direction from the distribution of all possible directions. This approach makes it possible to uncover connections missed by the deterministic algorithms (Fig. 6c) and has been proven to be superior to deterministic tracking. Consequently, modern tracking algorithms relying on advanced models, such as CSD, employ probabilistic tracking algorithms.

Even within the class of probabilistic algorithms, there exists a wide range of possible choices, which impacts the outcome tractogram. The difference between two probabilistic algorithm—iFOD2 (implemented in MRtrix) and parallel transport tractography (implemented in Trekker)—is demonstrated in Fig. 6c.

The term “seed” refers to locations chosen as starting points for reconstructed streamlines. Global seeding describes allowance to track from any brain voxel (usually white matter voxel), while ROI seeding refers to limiting seeds to defined ROIs. Global seeding allows for the reconstruction of all possible pathways, which is important in studies on the connectivity within the whole brain, but at the same time is much more computationally demanding. It also does not guarantee that the pathways of interest will indeed be reconstructed. The ROI seeding limits the tractography only to the selected structures, which is faster, but may, at the same time, limit the application of streamline post-processing options (see next section).

For the recommended probabilistic algorithms, the generated streamlines will not only be limited to the “true” pathways between seeds, but will also cover a wide range of positive, but anatomically implausible, connections (Fig. 6d). In order to limit the tractography outcome to only valid streamlines, it is advised to employ information about the destination of the streamlines, known as a “target.” A combination of information about start (seed) and end (target) ROIs greatly improves the accuracy of tractography (Fig. 6d).

Still, the definition of seed and target ROIs does not necessarily unambiguously determine the reconstructed streamlines, especially in the case of complex neural connections, such as the optic radiation. Therefore, it is necessary to always visually inspect the tractography results. For uncertain results, it is recommended to introduce further inclusion or exclusion ROIs (based on anatomical knowledge) in order to impose other restrictions on generated streamlines and remove implausible connections (Fig. 6d). A good and widely used example of the incorporation of anatomical priors in tractography is using the white matter mask as a limitation for the tractography. This option is supported in all tracking software, as well as its extensions to tissue types other than WM (such as in ACT; [34]).

Apart from the aforementioned points, it is also necessary to choose the parameters governing tracking, such as the number of steps, maximal angle between consecutive steps, and possible thresholds preventing tractography from entering false regions etc. A detailed discussion of this topic requires a separate study, as the selection of fixed parameters introduces tracking bias. A solution to this problem is the repetition of tractography multiple times (using different algorithms, sets of parameters, and possibly even models) and combining all these into one outcome [38]. This will, however, naturally extend the complexity of the analysis and its duration.

Once the tractogram is generated, it can be subjected to further editing and adjustments. All approaches can generally be grouped into two classes: (a) User-informed tractograms can be merged, divided with respect to the number of streamlines or their properties (such as length), split using newly defined ROIs, transformed to different templates etc. The common denominator here is the user-made decision about the action. (b) Signal-informed tractograms are automated methods that refine the selection of streamlines using initially measured signal as a reference, a validation process referred to as filtering. As an example, linear fascicle evaluation (LiFE; [29]) calculates the predicted signal based on generated streamlines and compares it with the currently measured signal. Redundant streamlines with zero contribution to the currently measured signal are subsequently discarded. Other filtering methods include, e.g., Spherical-deconvolution Informed Filtering of Tractograms (SIFT; [35]) or Convex Optimization Modeling for Microstructure Informed Tractography (COMMIT; [9]).