Measuring Connectivity in the Primary Visual Pathway in Human Albinism Using Diffusion Tensor Imaging and Tractography

Podsumowanie

This manuscript describes deterministic and probabilistic algorithms for white matter (WM) reconstruction, used to examine differences in optic radiation (OR) connectivity between albinism and controls. Although probabilistic tractography follows the true course of nerve fibers more closely, deterministic tractography was run to compare the reliability and reproducibility of both techniques.

Streszczenie

In albinism, the number of ipsilaterally projecting retinal ganglion cells (RGCs) is significantly reduced. The retina and optic chiasm have been proposed as candidate sites for misrouting. Since a correlation between the number of lateral geniculate nucleus (LGN) relay neurons and LGN size has been shown, and based on previously reported reductions in LGN volumes in human albinism, we suggest that fiber projections from LGN to the primary visual cortex (V1) are also reduced. Studying structural differences in the visual system of albinism can improve the understanding of the mechanism of misrouting and subsequent clinical applications. Diffusion data and tractography are useful for mapping the OR (optic radiation). This manuscript describes two algorithms for OR reconstruction in order to compare brain connectivity in albinism and controls.An MRI scanner with a 32-channel head coil was used to acquire structural scans. A T1-weighted 3D-MPRAGE sequence with 1 mm³ isotropic voxel size was used to generate high-resolution images for V1 segmentation. Multiple proton density (PD) weighted images were acquired coronally for right and left LGN localization. Diffusion tensor imaging (DTI) scans were acquired with 64 diffusion directions. Both deterministic and probabilistic tracking methods were run and compared, with LGN as the seed mask and V1 as the target mask. Though DTI provides relatively poor spatial resolution, and accurate delineation of OR may be challenging due to its low fiber density, tractography has been shown to be advantageous both in research and clinically. Tract based spatial statistics (TBSS) revealed areas of significantly reduced white matter integrity within the OR in patients with albinism compared to controls. Pairwise comparisons revealed a significant reduction in LGN to V1 connectivity in albinism compared to controls. Comparing both tracking algorithms revealed common findings, strengthening the reliability of the technique.

Wprowadzenie

Albinism is a genetic condition primarily characterized by overt hypopigmentation observed in affected individuals. It is caused by inherited mutations to genes involved in melanin synthesis¹. Albinism appears in two main forms: oculo-cutaneous albinism (OCA), an autosomal recessive trait presenting both ocular and cutaneous features; and ocular albinism (OA), an X-linked trait more prevalent in males and characterized primarily by the ocular symptoms². Melanin in the retinal pigment epithelium (RPE) is crucial for proper development of the central visual pathway. Its absence in albinism therefore results in visual impairments, including photophobia, nystagmus, reduced visual acuity and loss of binocular vision^2-3. Visual acuity has been linked to foveal morphology, which is altered in albinism⁴. In humans, a retinal line of decussation lies along the nasotemporal border through the fovea, with fibers from nasal retina crossing to the other hemisphere and those from temporal retina extending ipsilaterally. The degree of reduced visual function in albinism has been linked to the level of hypopigmentation. Specifically, pigmentation is inversely proportional to the shift into temporal retina of the line of decussation⁵. As a result of the shift in line of decussation into the temporal retina, crossing of optic nerve fibers is increased – a characteristic common across all species³.

Structural MRI studies on humans have shown narrower optic chiasms in albinism compared to controls, which is likely the result of increased crossing of RGCs observed in albinism^6-8. The retina and optic chiasm express axonal guidance cues such as Eph family receptors and their ligands⁹ and are therefore candidate sites for misrouting¹⁰.

A study on monkeys with induced glaucoma revealed a significant decrease in the number of LGN parvalbumin-immunoreactive relay neurons and LGN volume¹¹. This suggests a correlation between LGN size and the number of white matter (WM) trajectories traveling through the OR to V1. A post mortem study on human albinism also revealed smaller LGN with fused M and P layers¹². High-resolution structural MRI confirmed significant reduction in volume of LGN in albinism⁸. Taken together, these findings suggest that decreased LGN volume may result in a reduced number of neurons in the LGN, and in turn in decreased connectivity between LGN and V1.

Examining patterns of anatomical connectivity in humans has been limited. Dissection, tracer injection and lesion induction are invasive techniques that can only be used post mortem, and usually involve a very small number of patients. Previous studies using carbocyanine dye DiI injections demonstrated neuronal connectivity between V1 and V2 (secondary visual cortex)¹³, as well as within the hippocampal complex in aldehyde-fixed post-mortem human brains¹⁴. Labeling fibers in this way is restricted to distances of only tens of millimeters from the point of injection¹⁴. Diffusion tensor imaging, DTI, is an MRI modality developed in early-mid 1990s to identify fiber tract direction and organization. It is a non-invasive method that allows mapping of large WM pathways in the living brain. DTI is sensitive to the diffusion of water molecules in biological tissue¹⁵. In the brain, the diffusion of water is anisotropic (uneven) due to barriers such as membranes and myelin. WM has high diffusion anisotropy, meaning the diffusion is greater parallel to than perpendicular to the orientation of the fibers¹⁶. Fractional anisotropy (FA) is a scalar quantity that describes the preference of molecules to diffuse in an anisotropic manner. FA values range from 0-1, from low to high anisotropy (cerebrospinal fluid (CSF) <gray matter (GM) <WM)¹⁶.

Streamline (deterministic) and probabilistic fiber tracking are two different algorithms for 3D path reconstruction. Deterministic tractography uses a line propagation method, connecting neighboring voxels in a defined seed region. Two stop criteria used in this algorithm are the turning angle and the FA value. Therefore, tract tracing between neighbouring voxels is unlikely at large turning angles. The algorithm would therefore also progresses only if the FA in a voxel exceeds a specific threshold, limiting its effectiveness in accurately defining pathways near gray matter, where anisotropy drops. Probabilistic tractography, on the other hand, yields a connectivity map describing the probability of a voxel to be part of a tract between two regions of interest (ROIs) and thus progresses into gray matter such as V1¹⁷. Using this MRI application, key WM structures like the OR can be delineated, as shown in previous studies^18-20.

This study therefore uses diffusion data and tractography to explore the effect of axonal misrouting on retino-geniculo-cortical connectivity. Based on previously reported reductions in LGN volumes in human albinism⁸, we predict that fiber projections from LGN to V1 are also reduced (Figure 1).

Protokół

Ethics Statement: The current research study has been approved by the Human Participants Review Committee (HPRC) at York University, Toronto. All participants gave informed written consent.

1. Subject Preparation

Note: Eleven participants with OCA, aged 36 ± 4 years (6 females) were compared to ten age-matched controls, aged 32 ± 4 years (6 females). Participant history is recorded in Table 1.

1. Ask each participant to fill out and sign a consent form that lists MRI safety guidelines and imaging protocol.
2. For each participant, provide earplugs for the ears. Position participant supine and head first in the magnet, and landmark above the eyes at the eyebrow level. Secure participant's head with cushions to reduce head motion. Give the participant a squeeze bulb for patient alert.

2. Structural MRI Parameters

Note: All imaging is acquired on a 3T MRI scanner using a 32-channel head coil. During a single session per subject:

1. Acquire a high resolution T1-weighted anatomical using a 3D-MPRAGE sequence covering the entire brain with the following parameters: acquisition time 4 min 26 sec, field of view 256 mm, 256 matrix, 192 slices with slice thickness of 1 mm, with a resulting isotropic voxel size of 1.0 mm³, TR = 1900 ms, TE (echo time) = 2.52 ms with an inversion time of 900 ms and flip angle of 9°, 1 average, parallel imaging (iPat GRAPPA, acceleration factor of 2).
2. Acquire a DTI sequence covering the cortex, with slices in transverse orientation following the anterior commissure/posterior commissure (AC-PC) line, using the following parameters:acquisition time 8 min 5 sec, field of view 192 mm, 128 matrix, voxels 1.5 1.5 mm in-plane, 56 contiguous (no gap) slices with 2 mm thickness, TR = 6900 ms, TE = 86 ms, 64 directions, b-value of 1000 s/mm² (reference image with low b-value of 0 sec/mm²), 1 average, parallel imaging (iPat GRAPPA) with an acceleration factor of 3.
3. Acquire 30-40 PD-weighted images in a coronal orientation, parallel to the brain stem,covering from the anterior extent of the pons to the posterior portion of the inferior colliculus.
   1. Use the Turbo spin echo (FAST spin echo) pulse sequence and the following parameters:acquisition time 1 min 29 sec per scan, field of view 192 mm, 256 matrix, 30-40 slices with thickness of 1 mm, resulting voxel size 0.75 0.75 1 mm³, TR = 3,000 msec, TE = 22 msec, turbo factor of 5, refocusing flip angle of 120°, 1 average, parallel imaging (iPat GRAPPA) with an acceleration factor of 2.
      Note: S12 was scanned using the following parameters: field of view 180 mm, 512 matrix, 30 slices with 1 mm thick slices, resulting voxel size 0.4 x 0.4 x 1.0 mm³. All other parameters remained the same. Acquisition time 2 min 47 sec.
4. Pre-process all scans by converting raw DICOM to NIfTI format using the program dcm2nii.

3. LGN Delineation

Note: The LGN is a small subcortical structure located deep in the brain, hence high-resolution PD images are required to determine its anatomical boundaries. In these scans, the LGN appears as an area of high signal intensity relative to the surrounding WM tracts, facilitating its detection²¹. The identified anatomical LGN is then used as a seed region for tractography.

1. While blind to group membership, manually trace right and left LGN masks three times each on averaged PD images interpolated to twice the resolution and half the voxel size (original 256 x 256 matrix, 0.75 x 0.75 x 1 mm³ voxel size).
   1. To obtain high resolution PD images use the freely available FLIRT function and other software tools within FMRIB's Software Library (FSL, http://www.fmrib.ox.ac.uk/fsl/). Upsample, concatenate, motion correct and average PD images for each participant as previously described elsewhere²².
   2. Load the high resolution PD image in FSLView and click on the Tools tab to select the Single option (or press ) to enlarge the image.
   3. Click the File tab to select the Create Mask option, and use the toolbar on the top left of the screen to trace the LGN in each slice. If desired, change the contrast of the image by dragging along the min/max in the toolbar to facilitate LGN detection.
2. Merge these regions of interest (ROIs) into a median mask using the fslmerge command.
3. Combine all raters' median masks into a single median mask using the same command.

4. V1 Segmentation

1. Run "recon-all" command in FreeSurfer23 (v5.3.0) on brains in native anatomical space (T1-weighted images) for automated processing.
2. Convert the appropriate outputs in the newly created mri folder (orig.mgz, brain.mgz, rawavg.mgz, T1.mgz) to NIfTI using "mri_convert".
3. Use BET brain extraction in the FSL GUI to correct the skull-stripped output brain (brain.nii.gz) in FreeSurfer space if necessary. Choose the Run standard brain extraction using bet2 option (default). Lower the threshold if image is missing brain tissue, or increase if non-brain tissue is captured (default threshold 0.5). Select the Output brain-extracted image and Output binary brain mask image (the latter may be used for manual corrections) in the advanced options.
4. Convert output V1 parcellation to a volumetric mask using "label2surf" and "surf2volume" commands.

5. Pre-tracking Registrations

Note: For the next steps, call the FSL GUI to open each of the following tools.

1. Use BET brain extraction and select the Bias field & neck cleanup option to skull-strip rawavg.nii.gz, located in the mri folder created by "recon-all". Adjust threshold as necessary.
2. Run FLIRT linear registration to bring brains in FreeSurfer and native anatomical space to diffusion space.
   1. Select brain.nii.gz, output of recon-all (FreeSurfer space), or a subject's brain extracted T1 (native anatomical space) as the input image, and an Eddy corrected and brain extracted diffusion-weighted image (DWI) as the reference image. Then click "Go".
      Note: This step creates two outputs, the input brain registered to the reference image (.nii.gz) and a transformation matrix (.mat). Apart from registration, the latter file is required for tractography when seed space is not diffusion. Use the output transformation matrices (.mat) created in this step for tractography as explained in 7.4.2.
3. Similar to 5.2, run FLIRT linear registration to bring participants' PD brains to FreeSurfer space and native anatomical space.
4. Prepare seed masks for tractography:
   1. Apply FLIRT transformation from Utils in the FLIRT linear registration toolbox. Use the .mat output as the transformation matrix, the original LGN mask as the input and brain.nii.gz (FreeSurfer space) or T1_brain.nii.gz (native anatomical space) (see 5.2) as the reference volume. Select the Nearest Neighbor interpolation method from the advanced options.
5. Using brain.nii.gz files only, prepare target masks for tractography:
   1. Register FreeSurfer brains to native anatomical space and create target masks by applying transformation to V1 masks (see 5.2, 5.4.1) using Tri-Linear interpolation. Click "Go".

6. LGN Normalization 

1. Use FNIRT non-linear registration as described previously at http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FNIRT to bring participants’ non-extracted brains in native anatomical space to MNI space, using the Montreal Neurological Institute whole brain template (MNI152).
   Note: Non-linear registration of original anatomical images is recommended for this step, as registrations were more accurate when FNIRT was applied to non-extracted T1s compared to FLIRT on extracted brains.
2. Apply transformation to LGN masks in anatomical space (original LGN previously transformed to native anatomical space in 5.4) using nearest-neighbour interpolation as described in 5.4.1 to bring masks to MNI space.
3. Average all LGN masks in MNI space across both groups using AFNI's "3dMean" command.
4. Use "fslmaths -thr" to apply a threshold to the mean mask in MNI space.
5. Calculate the radius of the mean mask in MNI space using V = 4/3πr³ (assume a sphere).
6. Record the centre of mass coordinates of each individual LGN mask in native anatomical space using the command "fslstats -C".
7. Create spherical ROIs of identical volumes across participants:
   1. Use "fslmaths" to create an ROI point with the coordinates of the appropriate individual LGN mask in native anatomical space as recorded in 6.6
   2. Using "fslmaths", apply the radius of the mean mask in MNI space to create a sphere around the ROI point in native anatomical space.
8. Use these standardized masks as seeds for tractography.

7. Probabilistic Tractography (FSL 5.0.4)

Note: For the next steps, call the Fdt_gui to access each of the following tools.

1. Correct for distortions in DWIs with Eddy current correction. Select the Eddy current correction option from the menu at the top of the Diffusion Toolbox window and upload the DWI as the input, leaving the default reference volume (0).
2. Brain extract the images with BET as described in 4.3.
3. Select the DTIFIT Reconstruction diffusion tensors option from the menu. Specify an input directory containing the following files: diffusion weighted data, nodif_brain_mask (output of BET), bvec and bval (must be renamed to bvecs and bvals; text files containing information about diffusion image acquisition parameters, output of DICOM to NIfTI conversion of diffusion data). Click "Go" to run dtifit, which fits a diffusion tensor model at each voxel, creating files for post-processing.
4. Next, select the BedpostX (estimation of diffusion parameters) option from the menu. Use the same input directory as for DTIFIT. Click 'Go' to generate all files required for tractography.
5. From the same menu, choose ProbtrackX for probabilistic tracking and run it for each hemisphere separately. Keep default basic options (5,000 samples, 0.2 curvature and loopcheck applied) and select modified Euler for computing probabilistic streamlines from advanced options for increased accuracy.
   1. Select the output of BedpostX containing .merged files as the BEDPOSTX directory.
   2. Select single mask as seed space and load the transformed LGN mask (in native anatomical space) as the seed image, T1 (brain in native anatomical space) to diffusion transformation matrix as the seed to diffusion transform, and V1 (in native anatomical space) in "optional targets" (all but exclusion masks) as the target.
   3. Use default mesh convention (Caret) and load the brain in native anatomical space (T1 image) as the surface reference image.
6. Repeat ProbtrackX for probabilistic tracking using the standard spherical ROIs (created in step 6) as seed regions for tractography as described in 7.5.2. Upload ROIs in the same way transformed LGN (anatomical space) were uploaded in 7.5.2.
7. Re-run tractography (7.5), this time with seed (non-normalized) and target masks in FreeSurfer space with the addition of FreeSurfer's contralateral white matter border mask as an exclusion mask, to avoid any crossing over and ensure direct ipsilateral connections. Check the Surface option from the ProbtrackX toolbox and select FreeSurfer as mesh convention.
   Note: It is important to emphasize that tractography is always run from diffusion space, but Probtrackx for probabilistic tracking allows the input of seed and target masks in a different space, along with a transformation matrix to diffusion space. In this study, probabilistic tractography was run with masks in both native anatomical and FreeSurfer space (Figure 2).

8. Deterministic Tractography (DSI Studio)

1. Open Eddy corrected diffusion-weighted images in DSI Studio²⁴ by clicking on Step 1: Open Source Images. Load bvec and bval files onto a b-table window that is automatically opened to create a source (.src) file.
2. Load the generated Source files onto the reconstruction window to modify the default reconstructed brain masks as necessary.
3. Then, select DTI as the reconstruction method²⁵ and run it on the source files to produce fiber information files (.fib).
4. Bring Participants' PD brains to diffusion space using FLIRT linear registration.
5. Apply transformation to LGN masks using nearest neighbor interpolation as described in 5.4.1.
6. Open .fib files in the program's tracking window.
7. Run tracking for each hemisphere separately, using LGN in diffusion space as the seed and Region 17 (V1) from Brodmann atlas available from DSI Studio as the terminative region. Load the LGN mask by clicking the Regions tab and Open Region. Select the Seed option under Type in the Region List on the left of the screen. To load the V1 mask from the atlas, click on Atlas from the toolbar in the Region List and select the appropriate atlas.
8. In each run, set the contralateral WM (named left/right-cerebral-white-matter) mask from FreeSurfer segmentation atlas (see Region List box in the tracking window) as a region of avoidance (ROA).
9. Repeat tracking (8.7-8.8) using spherical ROIs in diffusion space instead of individual LGN as seed regions for tractography.
   Note: The spherical ROIs have the same volume across all subjects and are centered on the center of mass of each LGN.
10. Repeat LGN normalization, section 6, only this time registering brains in diffusion space to standard MNI space, and applying transformations to LGN in diffusion space (original LGN previously transformed to diffusion space in 8.4-8.5) to bring masks to standard MNI space. Calculate the volume of the spherical ROI as the mean volume of all LGN across subjects in MNI space.
    Note: Tracking parameters can be modified by the user. For most runs, default tracking parameters were applied. For some individuals (A5, A7, S12), anisotropy threshold (default 0.14-0.15) was lowered (0.10-0.12) and angular threshold (default 60) was increased (65-85) for nicer visualization. A schematic of the technique is shown in Figure 3.

9. Statistical Analysis – TBSS (FSL)

Note: Tract-based spatial statistics is a voxelwise statistical analysis of participants’ FA maps16 obtained with dtifit26. It is extensively used for statistics on diffusion data. This voxelwise approach overcomes potential alignment and smoothing problems seen in VBM-style FA analysis and provides whole brain investigation, unattainable through tractography-based approaches16.

1. Run "tbss_1_preproc" on the FA data located in a newly created TBSS directory.
2. Run "tbss_2_reg" – T to apply non-linear registration, bringing each participant's FA data into common space (FMRIB58_FA, target image in TBSS).
3. Create a mean FA skeleton with the centers of all common tracts among participants using "tbss_3_postreg -S".
4. Run "tbss_4_prestats 0.2" to project each participant's aligned FA map onto the mean skeleton of all aligned FA maps.
5. Create design.con and design.mat files, ensuring that the order of the matrix is consistent with the order in which TBSS pre-processed the FA data.
6. Run "randomise", using the T2 option, which is recommended for TBSS as it acts on a skeleton (a reduced subset of the 3D data), and 5,000 pre-mutations, which gives more accurate p-values.

10. Statistical Analysis – SPSS 

1. Extracting FA Values from Deterministic Data
   Note: Deterministic-based FA values were derived from DSI Studio output statistics text files. These values represent the mean FA within the generated tracts, which in this case correspond to the region of the OR.
   1. Run fiber tracking in DSI studio.
   2. Save the 'statistics' text files created by DSI Studio for each generated set of tracts and record the 'FA mean' values from them.
2. Extracting FA Values from Probabilistic Data
   Note: Probabilistic-based FA values are derived from ProbtrackX2 output fdt_paths files. These are 3D tract density images that in this study cover the area corresponding to the OR.
   1. Use FLIRT linear registration to bring each participant's fdt_paths files to diffusion space.
   2. Binarize the output masks using "fslmaths - bin".
   3. For each participant, multiply the mask by their FA map from dtifit using "fslmaths -mul".
   4. Run "fslmeants" command to find the mean FA from each tract mask.
3. Running Analyses with SPSS (Using Deterministic and Probabilistic
   Data)
   Note: Statistical analysis is performed using SPSS 20 for Mac. Since hemisphere is a within-subject variable, a generalized linear model (GENLIN) with which the effects in each side of the brain can be looked at separately, is applied. Specifically, the generalized estimating equation (GEE) is used.
   1. In separate tests, set each of mean FA and streamline count (waytotal or percentage generated streamlines, PGSL) as the dependent variable.
      Note: In this study, streamline count is based on way-total values. Waytotal describes the total number of generated streamlines that have not been rejected by inclusion/exclusion criteria²⁷. The number of generated streamlines (NGSL), which refers to the total number of streamlines sent, is equal to the number of voxels in the seed mask multiplied by the number of samples drawn from each voxel (5,000 in this case). Percentage generated streamlines (PGSL), waytotal divided by NGSL times 100, is a measure of successful connectivity between the seed and the target.
   2. Study the influence of group and gender on LGN to V1 connectivity by setting them as independent variables in all tests.
      Note: Main effects as well as two- and three-way interactions were studied. It is important to note that these individual tests are not conditioned to each other, so the significance of one main effect or interaction is independent of the other.
   3. Use age as a covariate for all tests. Also, use LGN volume as a covariate for tests with mean FA and waytotal as the dependent variables, but omit it from tests with PGSL as the dependent variable.
      Note: Total brain volume was found to be an insignificant covariate and was therefore omitted from stats.
   4. Select the Bonferroni correction method to adjust for multiple comparisons²⁸ (level of significance p <0.05).

Wyniki

This section provides a summary of results obtained using two different algorithms of tractography, deterministic and probabilistic. LGN volumes in PD space in which masks were originally drawn, as well as in all other spaces used in this study, are recorded in Table 2, and LGN tracing is illustrated in Figure 4. The results reported here are based on runs that used a standard sphere as the LGN ROI. Standard LGN volume was 461 mm³

Dyskusje

Altered WM and, more specifically, decreased connectivity in albinism compared to controls were expected. Thus, the reduced FA in the right hemisphere of albinism compared to controls as well as the decreased connectivity in male patients with albinism reported here are in line with our prediction. Gender and hemisphere effects are not completely clear, although research on the healthy brain that suggests decreased WM complexity in the left hemisphere of males compared to females^30-31 could explain some of the...

Materiały

Name	Company	Catalog Number	Comments
Magnetom Tim Trio 3T MRI	Siemens (Erlangen, Germany)
FMRIB’s Software Library (FSL)			http://www.fmrib.ox.ac.uk/fsl/
FreeSurfer			http://surfer.nmr.mgh.harvard.edu
DSI Studio			http://dsi-studio.labsolver.org
SPSS