-
Notifications
You must be signed in to change notification settings - Fork 0
Mahalanobis Distance (D2) in the Optic Radiations (Lab Rotation Project)
This is a Wiki of the methods used in a lab rotation project that examined variations of white matter microstructure in the optic radiations using Mahalanobis distance (D2) of DTI, NODDI and ODF scalars. The project also explored a possible link to differences in gender, age, visual acuity and contrast sensitivity.
There are relationships between brain structure and behavior as well as structural and functional variabilities across subjects. However, these studies mainly investigated variations in grey matter such as studies on cortical thickness and task performances. With regards to white matter (WM), a study reported that main trends of structural connectivity can predict task-related patterns of cortical activations (Thiebaut de Schotten and Forkel, 2020) which leads to the question how differences in white matter micorstructure relate to that finding.
Previous work in the lab focused on correlations of WM microstructure with behavioral variables such as grip strength, dexterity, processing speed, executive function (work of Zaki Alasmar and Stephanie Tremblay). For their research they used (diffusion)MRI and behavioral data of the Human Connectome Project (HCP). Moreover, a tool to quantify differences in WM microstructure was developed in the lab: the MVCOMP toolbox. It makes use of a measure called Mahalanobis distance (D2) which is defined as the multivariate distance between a point and a distribution in which covariance between features is accounted for. Thus, D2 can be used to calculate the difference in WM microstructure (as obtained by DTI, NODDI and ODF scalars) that a voxel of a specific subject has with respect to the average WM microstructure of this voxel across the rest of the study sample and accounting for covariance of the individual DTI, NODDI and ODF scalars.
For this project we decided to focus on WM microstructure and vision. The HCP dataset allowed the analysis of two visual features: visual acuity and contrast sensitivity. Visual acuity is defined as the sharpness and clarity of vision and the ability to discern fine details. Contrast sensitivity in turn is defined as the ability to detect differences in luminance (brightness) between an object and its background. While quite some evidence exists regarding the neural mechanisms of both features in the optics of the eye and the visual cortex, barely anything is known about the relationship of WM micorstructure of major fiber tracts for visual information transmission and visual behavior. The most important fiber tract for vision are the optic radiations (OR) that are connecting the lateral geniculate nucleus of the thalamus (LGN) with the primary visual cortex (V1). Decreases in fractional anisotropy (FA; a measure of white matter integrity) were associated with higher age (Kruper et al., 2022), poor visual acuity in preterm children (Thompson et al., 2014) and abnormal visual acuity in optic pathway gliomas (de Blank et al., 2013). No studies were found that report about a relationship between WM microstructure of the OR and contrast sensitivity measures.
This project aimed to examine differences in WM microstructure as assessed by D2 in the OR of subjects of the HCP dataset. Furthermore, it was tested in regression analyses if differences in WM microstructure relate to differences in sex, age as well as visual acuity and contrast sensitivity.
Data collection and sharing for this project was provided by the MGH-USC Human Connectome Project (HCP; Principal Investigators: Bruce Rosen, M.D., Ph.D., Arthur W. Toga, Ph.D., Van J. Weeden, MD). HCP funding was provided by the National Institute of Dental and Craniofacial Research (NIDCR), the National Institute of Mental Health (NIMH), and the National Institute of Neurological Disorders and Stroke (NINDS). HCP data are disseminated by the Laboratory of Neuro Imaging at the University of Southern California. The HCP dataset includes dMRI and behavioral data of 1001 subjects that were acquired according to a standard dMRI scan protocol and preprocessed following the Minimal Preprocessing Pipeline as specified by the HCP. For further details on participant selection and ethics please check the HCP website.
The data acquisition process of the HCP dMRI dataset involved a Siemens Connectom Skyra 3 Tesla scanner equipped with a 32-channel head coil. The imaging parameters included an echo time (TE) of 89.5 ms and a repetition time (TR) of 5520 ms. The field of view (FOV) was set to 210×180 mm, ensuring comprehensive coverage of the brain. To capture a wide range of diffusion-weighted information, a multi-shell approach was employed, consisting of b-values of 1000, 2000, and 3000 s/mm². The imaging resolution was set to 1.25 mm isotropic, providing high spatial fidelity. To adequately sample diffusion orientations, 90 uniformly distributed directions were acquired. Additionally, six volumes were acquired with a b-value of 0 to serve as reference images for subsequent processing and analysis.
The minimal preprocessing pipeline applied several essential preprocessing steps to the acquired data. Firstly, intensity normalization was performed on the b0 images to ensure consistent signal levels. Subsequently, correction methods were applied to address eddy current and susceptibility-induced distortions, as well as motion and gradient nonlinearity artifacts. The resulting images were then registered to the native structural space, utilizing a rigid transform and aligning them with the T1-weighted image. Tensor calculations were performed, allowing the derivation of diffusion tensor imaging (DTI) metrics such as fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD). To account for inter-participant variability, response functions for each tissue type were estimated and averaged across participants. Furthermore, orientation distribution functions (ODFs) were computed for each tissue type using a multi-shell multi-tissue constrained spherical deconvolution (CSD) approach. Finally, a bias field correction was applied to mitigate intensity variations across the image, followed by global intensity normalization to ensure consistency and comparability across the dataset. For further details on the Minimal Preprocessing Pipeline for dMRI data, please check the HCP GitHub page.
The analyses were carried out using Python and some minor linux bash scripting for generating the OR mask. The following toolboxes and packages were used: MVCOMP (developed in the lab), MRTrix 3, FSL, ANTs, pandas, numpy, nibabel, sklearn, matplotlib, seaborn, nilearn, scipy, statsmodels.
To create the mask of the OR fiber tract, it was decided to perform whole-brain tractography using tckgen of MRTrix 3 on the averaged fod maps that have been previously created in the lab using the preprocessed data. The resulting streamlines were filtered for tracts starting at the LGN and ending at V1 using tckedit of MRTrix 3. The atlas used for defining LGN and V1 is HCPex (Huang et al., 2022), which was registered to the group space of the DWI data. To exclude unreasonable streamlines, an exclusion mask was manually defined (using the ROI tool in mrview) for streamlines running through midline, anterior, ventral and dorsal brain regions, as well as the posterior commissure. Using tckmap of MRTrix 3 the resulting streamlines were projected to volumetric space. The resulting OR mask was masked with a previously generated 99% WM mask of the HCP subjects that only includes voxels that belong to white matter in 99% of the HCP subjects. This was done to reduce partial volume effects that would bias the D2 calculation.
D2 was calculated using model_comp of the MVCOMP toolbox that has previously been generated in the lab. The distribution of the resulting D2 values were inspected for outliers. A subject was excluded from the study sample if 0.1% of the voxels had an unreasonably high d2-value of 5 or higher. For this project that was the case in n = 38 subjects which were excluded from the study sample. Based on the new sample of n = 963 subjects, D2 was recomputed. To make the D2 values more comparable to the behavioral scores, the D2 data was power transformed and z-standardized.
Age was given as positive integers for age of the subjects in years. The raw sex variable used "M" for male and "F" for female participants. Sex was recoded so that 0 was used for male and 1 for female sex. All variables were z-transformed and outliers with a z-score greater or less than 3 and -3 were excluded.
The HCP dataset provided behavioral data of two visual features: visual acuity and contrast sensitivity. Subjects with missing values of either variable were excluded from the study sample.
Visual acuity was assessed using the Electronic Visual Acuity Test (EVA). The test involves presenting a series of standardized optotypes, in this case letters, on a digital screen at varying sizes and asking the person to identify them. By determining the smallest optotype that can be recognized accurately, the test provides an objective measurement of visual acuity and aids in diagnosing and monitoring conditions affecting vision, such as nearsightedness or farsightedness. For the analyses, the EVA denominator was used which is defined as the distance in feet that a person with normal vision can read letters as well as the subject can read letters at a virtual distance of 20 feet (with corrected vision, if applicable). Thus, a value greater than 20 means the subjects has poorer than normal vision, while a value less than 20 means the subject has better than normal vision. For easier interpretation the variable was z transformed and multiplied by -1, so that positive EVA denominator values mean better than mean of sample visual acuity.
Contrast sensitivity scores were obtained by using the Mars Contrast Sensitivity Test (Mars). The Mars contrast sensitivity test is designed to assess an individual's ability to distinguish between light and dark contrasts. The test typically involves presenting a series of patterns or shapes, in this case letters, with varying levels of contrast on a computer screen and asking the person to identify them. The test continues until two consecutive errors have been made. In this project, the final score was used which is the value of the final correct letter, minus 0.04 for each incorrect letter prior to the two final consecutive errors. For both variables outliers were excluded for z-values greater or less than 3 and -3, respectively.
For the regression analyses nilearn.mass_univariate.permuted_ols was used. The D2 values were included as the target_vars and the respective explanatory variables and confounding variables as tested_vars and confounding_vars, respectively. N = 10000 permutations and threshold-free cluster enhancement (TFCE) were used. Family-wise error correction (FWER) is performed by default in permuted_ols. To estimate the effects of age and sex on D2 values two separate regression analyses were set up without confounding variables. With regards to behavior, each relationship was tested using age and sex as confounding variables.
The final data sample included N = 938 subjects with an age range from 22 to 36 years (mean = 28.8, std = 3.7). The sample consisted of slightly more females than males (56.2%). Interestingly, younger participants included more males, while older subjects included more females. Only N = 19 subjects had a kind of color blindness and only N = 67 subjects used some sort of vision correction.
After exclusion of outliers and subjects with missing values the study sample had a raw mean EVA denominator score of 15.08 (std = 4.73) and a mean Mars final score of 1.80 (std = 0.07).
The D2 scores varied from 0 to 5. Averaged across subjects some clusters of higher D2 were found in the right middle OR.
The permuted OLS regression of D2 values and sex revealed different WM microstructure in right anterior OR for different sex (p < 0.05, TFCE-based FWER corrected). The relationship was negative meaning that males tend to have higher D2 scores, so a greater difference in WM microstructure, in the right anterior OR than females (t-values around -3.96). The permuted OLS regression of D2 values and age did not show any significant voxels that survived multiple comparison correction. However, there were a few positive and negative t-values of around 3.6 and -3.6 that were signifiant for p < 0.05 (uncorrected).
The permuted OLS regressions for both visual acuity scores and contrast sensitivity scores did not show any significant voxels that survived multiple comparison correction.
This lab rotation project examined differences in WM microstructure as assessed by D2 in the OR using dMRI and behavioral data of the HCP project. Regional differences in D2 scores were observed with a cluster of higher D2 scores in the right middle OR. The only results with t-values surviving multiple comparison correction was a negative relationship of sex and D2 values in the sense that males tend to have a different WM microstructure in the right anterior OR.
xxx
Avidan G, Harel M, Hendler T, Ben-Bashat D, Zohary E, Malach R. Contrast sensitivity in human visual areas and its relationship to object recognition. J Neurophysiol. 2002 Jun;87(6):3102-16. doi: 10.1152/jn.2002.87.6.3102. PMID: 12037211.
de Blank PM, Berman JI, Liu GT, Roberts TP, Fisher MJ. Fractional anisotropy of the optic radiations is associated with visual acuity loss in optic pathway gliomas of neurofibromatosis type 1. Neuro Oncol. 2013 Aug;15(8):1088-95. doi: 10.1093/neuonc/not068. Epub 2013 May 7. PMID: 23658320; PMCID: PMC3714157.
Groppo M, Ricci D, Bassi L, Merchant N, Doria V, Arichi T, Allsop JM, Ramenghi L, Fox MJ, Cowan FM, Counsell SJ, Edwards AD. Development of the optic radiations and visual function after premature birth. Cortex. 2014 Jul;56:30-7. doi: 10.1016/j.cortex.2012.02.008. Epub 2012 Mar 8. PMID: 22482694.
Huang CC, Rolls ET, Feng J, Lin CP. An extended Human Connectome Project multimodal parcellation atlas of the human cortex and subcortical areas. Brain Struct Funct. 2022 Apr;227(3):763-778. doi: 10.1007/s00429-021-02421-6. Epub 2021 Nov 17. PMID: 34791508.
Kruper J, Benson NC, Caffarra S, Owen J, Wu Y, Lee AY, Lee CS, Yeatman JD, Rokem A; UK Biobank Eye and Vision Consortium. Optic radiations representing different eccentricities age differently. Hum Brain Mapp. 2023 Jun 1;44(8):3123-3135. doi: 10.1002/hbm.26267. Epub 2023 Mar 10. PMID: 36896869; PMCID: PMC10171550.
Tamada T, Enatsu R, Saito T, Chiba R, Kanno A, Mikuni N. Visual networks: Electric brain stimulation and diffusion tensor imaging. Rev Neurol (Paris). 2023 Apr 21:S0035-3787(23)00918-9. doi: 10.1016/j.neurol.2022.12.011. Epub ahead of print. PMID: 37088608.
Thiebaut de Schotten M, Forkel SJ. The emergent properties of the connected brain. Science. 2022 Nov 4;378(6619):505-510. doi: 10.1126/science.abq2591. Epub 2022 Nov 3. PMID: 36378968.
Thompson DK, Thai D, Kelly CE, Leemans A, Tournier JD, Kean MJ, Lee KJ, Inder TE, Doyle LW, Anderson PJ, Hunt RW. Alterations in the optic radiations of very preterm children-Perinatal predictors and relationships with visual outcomes. Neuroimage Clin. 2013 Nov 28;4:145-53. doi: 10.1016/j.nicl.2013.11.007. PMID: 24371797; PMCID: PMC3871291.
Yu H, Xu F, Hu X, Tu Y, Zhang Q, Ye Z, Hua T. Mechanisms of Surround Suppression Effect on the Contrast Sensitivity of V1 Neurons in Cats. Neural Plast. 2022 Mar 8;2022:5677655. doi: 10.1155/2022/5677655. PMID: 35299618; PMCID: PMC8923783.