Computational protocol: Using High Angular Resolution Diffusion Imaging Data to Discriminate Cortical Regions

Similar protocols

Protocol publication

[…] The GM/WM and pial surfaces were identified in FreeSurfer , . The test/re–test DWIs were first aligned to each other on a per diffusion encoding direction basis in AFNI ( (using 3dvolreg, heptic interpolation) to correct for discrepancies between the datasets. Pooled across the subjects, the maximum ranges of the amount of movement correction between the two datasets were well under a degree for rotation and under a millimeter for translation, except along the phase–encoding direction where apparent motion, due to Bo drift, was observed. Since, the professional volunteers were highly compliant, as demonstrated by the minimal amount of correction required across data sets, the only additional within–data–set correction applied was the monotonic image translation to cancel the B0 drift along the phase encoding direction. Specifically, the measured translation, T, between successive reference (b = 100 s/mm2) image volumes at the beginning of the test and retest datasets was used move the above aligned test–retest DWIs incrementally (i.e. 1*T/68 for the first image volume, 2*T/68 for the second image volume, etc).The first reference DWI image was then aligned with quantitative T1 map images using manual blink comparison (contrast–inverting in–house version of FreeSurfer tkregister). The resulting 4×4, affine, transformation matrix was used to align the 61 directions of Bo drift–corrected HARDI data with the higher resolution GM/WM surface reconstruction (approximately 150,000 vertices per hemisphere). From each vertex point on the GM/WM boundary surface, the direction of the local normal vector was followed to half way between the GM/WM boundary and the pial surface. At that point a single sample of signal intensity was taken from each, aligned DWI. The diffusion data for each direction was smoothly interpolated (within–direction) onto the higher resolution surface mesh using iterative nearest neighbor estimation . The estimated FWHM was a 1.8 mm surface kernel, which was smaller than the resolution of the DWI images. This process resulted in 61 data points per acquisition (representing the 61 DWIs) at each GM/WM surface vertex point, which were written out to separate files that could be used to visually check the diffusion data on the folded and unfolded surface and for export to Matlab 7.10 (MathWorks, Natick, USA) for further processing. At each surface vertex, we saved the unique vertex ID, the x, y, z coordinates of the vertex, the x, y, z components of the local normal vector (n), a unique voxel ID (i.e. because the anatomical images had a higher spatial resolution than the DWIs multiple surface vertices may sample a single, coarser DWI voxel), and finally, the 61 image intensity values extracted from the DWIs. [...] A spherical harmonic model was fit to the log HARDI data to obtain the apparent diffusion coefficient profile f as in . This model includes spherical harmonic terms up to the 6th order from which seven types of orientationally invariant features of the HARDI profile were computed. Features 1, 2 and 7 are independent of the local normal and fully orientationally invariant features, 3–6 are relative to the local normal, n. Specifically,The mean of the ADC profile(1)The kth moment of f for k = 2.10(2)where S is the unit sphere.The value of f along n to the local cortical surface(3)The mean of f perpendicular to the local n, (i.e. the mean ADC in the plane of the cortex)(4)where C(n) is the unit circle perpendicular to n.The kth moment of f perpendicular to n, for k = 2…10(5)The two eigenvalues of the Hessian matrix of f evaluated at n. The Hessian matrix is the second derivative of the ADC profile, which expresses its curvature and is sensitive to the dispersion of fibre orientations within the tissue .Simple rotationally invariant combinations of the spherical harmonic parameters for k = 0, 2, 4, 6(6)where aki is the coefficient of the spherical harmonic order k and index i in the series.In total this provides a feature vector of 27 values for every vertex point, which were used to differentiate distinct cortical areas. A principal component analysis suggested that the data actually contained around 9 or 10 significant degrees of freedom. However, to test the discriminative potential of the features, the full feature vectors were used as input to an off–the–shelf support–vector machine (SVM) classifier ( Experiment 1, data from the first acquisition were used to train the three–way SVM classifier on the full set of feature vectors from every vertex point within three regions. In Subject 1 (male) these three regions were MT+ (Extended middle temporal area based on retinotopy and quantitative T1 data ), Ang (a nearby region of the angular gyrus within the so–called “default mode network” that is known to be lightly myelinated ) and STS ROI (a visually responsive part of the superior temporal sulcus), as displayed in . An additional region, just anterior to MT+, was not included in training the classifier on data from the 1st acquisition but data from that region from the 2nd repetition was subsequently classified to investigate whether the classifier would find borders of regions automatically. In Subjects 2 and 3 (1 male/1 female) three regions were defined solely on quantitative T1 data . The regions used for the 3–way classification were Ang (as above for Subject 1), the region anterior to MT+ and M–I (primary motor cortex).The classification rate of the classifier on the second, unseen, acquisition provides an indication of discriminability of different cortical areas based on the HARDI signal. The classification results were also painted on the cortical surface using FreeSurfer for visual representation.In order to test the method on a larger number of distinct regions, in Experiment 2 data were extracted from additional areas. For Subject 1, nine areas were chosen as in using a combination of independent anatomical and functional criteria: quantitative T1 for primary sensory/motor areas and retinotopic functional imaging data for the remaining areas. Namely, A–I,R = primary auditory cortex and rostral area; FST = fundus of the superior temporal sulcus area; IPS1 = Lateral intraparietal sulcus area 1; IPS2,3 = lateral intraparietal sulcus areas 2 & 3; S–I = primary somatosensory cortex (areas 3b,1,2); V1 = primary visual cortex; V3A = V3 Accessory; V6 = visual area 6 and VIP = ventral intraparietal area. For subjects 2 and 3, seven regions were defined in addition to the Ang, region anterior to MT+ and M–I solely on quantitative T1 data. These were MT+, A–I,R, S–I, V1, V3A, V6 and VIP. The MT+ and the region anterior to it were included to specifically test the discriminability of adjacent regions. The ability of the SVM to distinguish the regions was tested pair–wise on data from these twelve regions in Subject 1 or ten regions for Subjects 2,3. For each pair of regions the SVM was trained on data from repetition 1 and then data from the same two regions in repetition 2 were classified. Each pair–wise classification produces two results: the fraction of correctly classified voxels in each of the two regions. We average these two classification rates to get an overall classification rate for each pair. This avoids misleading high scores from unbalanced pairs of regions with very different sizes. […]

Pipeline specifications

Software tools AFNI, LIBSVM, FreeSurfer
Applications Miscellaneous, Functional magnetic resonance imaging
Organisms Homo sapiens