Supplemental Data

Can fully automated detection of corticospinal tract damage be used in stroke patients?

Nancy Kou, MSc.1,2, Chang-hyun Park, PhD 1, Mohamed L. Seghier, PhD 3, Alexander P. Leff, PhD 2,4, Nick S. Ward, MD 1,4

1. Sobell Department of Motor Neuroscience, UCL Institute of Neurology, 33 Queen Square, London WC1N.

2. Institute of Cognitive Neuroscience, 17 Queen Square, London WC1N 3BG.

3. Wellcome Trust Centre for Neuroimaging, UCL Institute of Neurology, 12 Queen Square, London, WC1N 3BG.

4. The National Hospital for Neurology and Neurosurgery, Queen Square, London WC1N 3BG.

ONLINE SUPPLEMENTAL DATA

e-Methods

1) Motor performance assessments

The Action Research Arm Test (ARAT), Nine-Hole Peg Test (NHPT), Motricity Index (MI), and grip strengthwere administered.

The ARAT assesses pinch, grasp, grip, elbow flexion and extension, and gross arm movements for an overall characterization of activities in coordination, and dexterity. Performance of each movement is rated on a 4-point ordinal scale with a maximum score of 57 points.

The NHPT measured the time taken to place nine pegs with each hand. If patients failed to place all nine pegs within 60 s, the number of pegs successfully placed was recorded. Scores were recorded as pegs per second for each hand (averaged over three trials). The score for the impaired hand was corrected within subject by dividing by the score for the unimpaired hand.
The MI was used to assess patients’ upper limb ability to execute a range of motion, to resist applied force, and to move a specific muscle group. Performances of three tests (pinch grip, elbow flexion, and shoulder abduction) are indicated with scores ranging from 0 to 33 for each element. These three scores are summed and +1 added for a total score of 100.

The maximum grip strength of each hand was measured (in kgs) with the JamarPlus Digital Hand Dynamometer. The maximum value of three trials was taken for each hand and the ratio of affected over unaffected hand maximum voluntary grip force was calculated.

2) MRI protocols

All scans were performed in one single scanning session with a 3.0 Tesla Siemens Allegra MRI scanner (Erlangen, Germany). Structural and diffusion images were collected from all controls, and structural images were collected from patients. T1-weighted structural images were acquired in 176 slices (224 x 256 in-plane matrix; 1mm3 isotropic voxels). Sixty-eight diffusion-weighted images were acquired in 60 slices (96 x 96 in-plane matrix; 2.3mm3 isotropic voxels), 61 of which are high diffusion-weighted (b=1000s/mm2) and additional 7 have minimal diffusion weighting (b=100s/mm2) for tensor modeling.

3) Image preprocessing, probabilistic tractography

All four cortical seed ROIs were determined based on the Harvard-Oxford atlas, which defines the precentralgyrus and SMA. From the precentralgyrus and in MNI coordinates, M1 was demarcated to be the posterior half above z = 40mm, PMd to the anterior half above z = 40mm, and PMv to the anterior half below z = 40mm. Cortical masks were transformed into each control subject’s diffusion space. The cerebellum was also demarcated from the atlas and transformed into individual diffusion space. Waypoint regions were identified manually using FSLview ( at the level of the posterior limb of the internal capsule, ventral upper pons, and ventral lower pons in each subject’s native diffusion space. A mask at the level of the corpus collosum was also manually defined, and in combination with a cerebellum mask formed an exclusion mask.

For tractography, using FMRIB’s Diffusion Toolbox FDT v2.0 ( diffusion images were first realigned to the first image to correct for motion artefacts. Brain Extraction Tool (BET)e1 was then applied to isolate brain tissue. Fitting of diffusion tensors on each voxel was estimated using DTIFIT. Further, BEDPOSTX was performed to apply the probabilistic diffusion model with a 0.2 threshold and at 3 directions per voxel. Tractography in each subject’s native diffusion space was performed using the probabilistic tracking algorithm of FDT. 5000 numbers of fibres were modelled per voxel. Finally, eight tracts (four tracts from the corresponding cortical seed masks in each hemisphere) were acquired from each control subject. All tracts were transformed to the standard MNI space for subsequent overlaps across subjects.

Two types of template corticospinal tracts (CST) were calculated: binary and weighted. To eliminate the less probable section in a tract, each tract was assigned a weighting factor or threshold at 5% of the maximum probability and then binarised (i.e. a designated value of 1 corresponding to a voxel above the threshold, and a value of 0 indicating a voxel below the threshold). Each control subject’s binarised tract was then superimposed across our 23 control subjects to yield a voxel-wise overlap of the tract at the group level. The group-level tract was then thresholded at approximately half of the number of control subjects (12) and binarised again. All binarisation were omitted when calculating a weighted tract.

4) Lesion definition

The manual tracing of the lesion was performed on the normalized T1-weighted structural images from each patient in FSLview. A binary mask that covered the whole extent of the lesion was created and confirmed by a consultant neurologist.

A fully automated lesion identification (ALI)e2was implemented with New Segment in SPM8,e3 which increases the number of tissue classes to include bone, soft tissue, and air/background. The additional tissue classes would enhance the specificity of the outlier detection algorithm of ALI. The segmented grey and white matter tissues underwent smoothing at 8mm full-width-half-max (FWHM), and smoothed tissues of each patient were compared to those of healthy controls using a non-iterative fuzzy clustering with fixed prototypes.e4 The fuzzy clustering thus estimates for each patient a degree of abnormality at each voxel being an outlier (i.e. part of the infarct). Ventricular dilations and other voxel abnormalities (i.e. caused by small vessel disease) were also highlighted. Within this map, the degree of abnormality values were read and averaged. Both the binary and weighted lesion masks were transformed into standard MNI space.

5) Lesion-tract overlap

Four combinations of overlaps between the binary and weighted tracts, and binary and weighted lesion masks were performed. The overlap output was calculated by multiplying the template CST image by the lesion mask voxel-by-voxel.
Different weightings were applied to the cross-sectional area of each slice containing the reconstructed CST as it descends from cortical motor regions through the internal capsule into the brainstem.

Statistics

A principal component (PC) analysise5 of the behavioural scores (Action Research Arm Test (ARAT), the Nine-Hole Peg Test (NHPT), Motricity Index (MI), and grip strength) was conducted. The relationships between each of the four overlap methods and PC of behavioural scores with template CST from all cortical motor areas were assessed using the GLM in SPSS 20.0 (IBM, New York, US) (Table e-2).e6 A Bonferroni multiple comparisons correction was applied at a p-value of 0.05. Fisher’s r-z transformations between the Pearson’s correlation coefficients of the two methods were also performed (Table e-2). Intraclass Correlation Coefficients (ICC)e7 were calculated in SPSS using a two-way mixed model to illustrate the consistency between the lesion overlap methods (Table e-3).

e-Results

There was no difference in age between patients and controls (p = 0.330). The significance of all four overlap methods demonstrates their equivalence (Table e-2). The two selected methods (binary tract and binary lesion, and binary tract and weighted lesion) are thus representative of the four methods and are subsequently correlated with PC of behavioural scores. Different weightings applied to each cross-sectional area of the CST did not yield any significance in association with the global measure of motor impairment.

Table e-1

Patient ID / Age (years) / Time since stroke (months) / Gender / Side affected / Lesion Location / Lesion (voxels) / Motor Performance
ARAT
(max 57) / GRIP
(A/U %) / MI
(max 100) / NHPT
(A/U %)
1 / 69 / 30 / F / L / SC / 1216 / 56 / 84 / 85 / 19.8
2 / 69 / 38 / F / L / SC / 1694 / 57 / 94 / 93 / 50
3 / 77 / 26 / F / R / SC / 397 / 38 / 57.2 / 77 / 9
4 / 52 / 24 / M / L / SC / 193 / 42 / 35 / 88 / 9
5 / 62 / 38 / M / L / SC / 555 / 36 / 31 / 72 / 8
6 / 46 / 8 / M / R / SC / 451 / 57 / 106.2 / 100 / 104.5
7 / 53 / 31 / F / L / SC / 86 / 50 / 40 / 91 / 50
8 / 63 / 8 / M / R / SC / 44 / 57 / 91 / 100 / 77.9
9 / 58 / 11 / M / L / SC / 4 / 57 / 88.2 / 100 / 87
10 / 51 / 60 / M / R / SC / 380 / 45 / 104 / 92 / 31
11 / 45 / 29 / F / L / SC / 2484 / 45 / 51 / 85 / 35
12 / 58 / 13 / M / L / SC / 11267 / 37 / 66 / 64 / 9
13 / 42 / 26 / M / L / SC / 8600 / 44 / 80.6 / 85 / 0
14 / 70 / 3 / M / L / SC / 26025 / 57 / 81.3 / 100 / 81.3
15 / 69 / 9 / M / R / SC / 1761 / 57 / 80.5 / 100 / 69.7
16 / 55 / 5 / F / L / SC / 176 / 55 / 64 / 93 / 97
17 / 46 / 26 / M / L / SC / 6220 / 46 / 52.3 / 84 / 38
18 / 61 / 13 / M / L / SC / 914 / 45 / 51.1 / 65 / 19.7
19 / 75 / 6 / M / L / SC / 549 / 57 / 96.6 / 100 / 73.7
20 / 66 / 5 / M / L / SC / 86 / 57 / 63.4 / 92.5 / 98.2
21 / 44 / 8 / M / L / SC / 3554 / 36 / 78.6 / 81 / 5.1
22 / 36 / 20 / M / R / SC / 146 / 54 / 81.9 / 93 / 31
23 / 59 / 165 / F / L / SC / 4399 / 29 / 18.2 / 68 / 0
24 / 66 / 76 / M / L / SC / 2232 / 21 / 46.6 / 72 / 0
25 / 43 / 20 / F / R / SC / 6067 / 41 / 71 / 91 / 31
26 / 38 / 4 / M / R / SC / 157 / 57 / 71.7 / 100 / 68.9
27 / 50 / 13 / M / L / SC / 27712 / 31 / 35.3 / 42 / 0
28 / 40 / 21 / M / R / SC / 2406 / 48 / 52 / 91 / 53
29 / 40 / 25 / M / L / SC / 272 / 26 / 56.3 / 74 / 4
30 / 48 / 7 / M / L / SC / 3697 / 54 / 87.6 / 88 / 40.8
31 / 57 / 17 / M / R / SC / 1014 / 37 / 107 / 91 / 40
32 / 56 / 13 / M / R / SC / 1035 / 57 / 93.1 / 100 / 94.5
33 / 18 / 5 / M / L / SC / 696 / 26 / 24.9 / 48 / 5.1
34 / 22 / 3 / M / R / SC / 82 / 57 / 64.2 / 100 / 89.6
35 / 51 / 4 / M / L / SC / 1 / 19 / 97.7 / 100 / 51.3
36 / 52 / 83 / M / L / SC / 36 / 57 / 93.3 / 100 / 74.5
37 / 55 / 9 / M / L / SC / 36 / 56 / 31.8 / 85 / 14.7
38 / 60 / 41 / M / L / C / 13313 / 39 / 20.1 / 65 / 0
39 / 59 / 79 / M / L / C / 17489 / 21 / 50.3 / 73 / 0
40 / 59 / 3 / M / L / C / 61691 / 52 / 64.4 / 42 / 14.9
41 / 66 / 26 / M / R / C / 9566 / 35 / 81 / 65 / 39
42 / 58 / 7 / F / L / C / 10 / 52 / 111.7 / 93 / 60.6
43 / 62 / 3 / F / R / C / 9459 / 36 / 44 / 77 / 9
44 / 33 / 63 / F / R / C / 9021 / 39 / 41 / 77 / 9
45 / 56 / 16 / M / R / C / 22079 / 57 / 68 / 100 / 60
46 / 53 / 12 / F / L / C / 43496 / 55 / 60 / 93 / 30
47 / 80 / 20 / F / L / C / 33005 / 45 / 65 / 91 / 21
48 / 48 / 116 / M / R / C / 9441 / 57 / 84 / 84 / 71
49 / 42 / 22 / F / L / C / 74559 / 28 / 41.3 / 66 / 8.2
50 / 50 / 18 / F / L / C / 51851 / 44 / 27.7 / 93 / 5.9
51 / 60 / 6 / F / R / C / 39594 / 27 / 31 / 73 / 0
Mean / 53.9 + 12.7
Control / 50.6±14.7
3-165 / 16 F / 18 R / 14C / 1 -
74559 / 19-57 / 18.2-111.7 / 42-100 / 0-104.5

Overview: patient group characteristics.

The gender, side affected, age (years), lesion location, manually defined lesion volume (voxels), time since stroke (months), and motor impairment scores (ARAT, GRIP, MI, NHPT) of each patient are described.
‘SC’ indicates a lesion leaving cortical motor areas spared. These patients had infarcts in the striatocapsular, pontine, corona radiata, thalamus, and cerebellar regions due to occlusions in the anterior choroidal artery and branches of the middle cerebral artery.
‘C’ indicates a cortical lesion affecting cortical motor areas. Other non-motoric cortical regions may also be impaired.
F = females; M = males; L = left; R = right; UL = upper limb; A = affected; U = unaffected

Table e-2

Cortical area / Lesion-overlap with binary tract
Manual
(binary tract-binary lesion) / Automated
(binary tract-weighted lesion) / Fisher’s transformation of
parametric GLM
r
p / R2 / r
p / R2 / (z-score)
p
M1 / 0.46
0.001 / 0.21 / 0.56
<0.001 / 0.31 / (-0.66)
0.509
PMd / 0.43
0.002 / 0.18 / 0.57
<0.001 / 0.32 / (-0.90)
0.368
PMv / 0.32
0.022 / 0.10 / 0.45
0.001 / 0.21 / (-0.77)
0.441
SMA / 0.34
0.010 / 0.11 / 0.42
0.002 / 0.18 / (-0.50)
0.617
Cortical area / Lesion-overlap with weighted tract
Manual
(weighted tract-binary lesion) / Automated
(weighted tract-weighted lesion) / Fisher’s transformation of
parametric GLM
r
p / R2 / r
p / R2 / (z-score)
p
M1 / 0.48
<0.001 / 0.24 / 0.55
<0.001 / 0.30 / (-0.41)
0.682
PMd / 0.44
0.001 / 0.20 / 0.56
<0.001 / 0.31 / (-0.75)
0.453
PMv / 0.34
0.016 / 0.11 / 0.46
0.001 / 0.21 / (-0.71)
0.478
SMA / 0.36
0.011 / 0.13 / 0.42
0.002 / 0.18 / (-0.40)
0.689

Automated and manual lesion overlap with template corticospinal tract from all cortical motor regions. The two selected methods (binary tract and binary lesion, and binary tract and weighted lesion) are representative of the four methods and are used in subsequent correlations with PC of behavioural scores. No significant differences between correlation coefficients for any of the four corticospinal tract pairs were observed.

Table e-3

Cortical
area / ICC
p<0.0125 / CI of ICC
(95%) / Pearson’s (interclass) correlation (parametric)
(p <0.0125)
M1 / 0.71 / 0.57-0.83 / 0.73
PMd / 0.74 / 0.58-0.84 / 0.75
PMv / 0.79 / 0.63-0.86 / 0.80
SMA / 0.80 / 0.68-0.88 / 0.83

Comparison of Two-Way Mixed-Model Intraclass Correlation Coefficient (ICC)and Pearson’s correlation coefficients from all cortical motor regions between the automated and manual lesion identification methods.

e-REFERENCES

e1.Smith SM. Fast robust automated brain extraction. Hum Brain Mapping 2002;17:143-55.

e2.Seghier ML, Ramlackhansingh A, Crinion J, Leff AP, Price CJ. Lesion identification using unified segmentation-normalisation models and fuzzy clustering. Neuroimage 2008;41:1253-1266.

e3.The FIL Methods Group, 2010. SPM8 Manual. manual.pdf. 2010.

e4.Seghier ML, Friston KJ, Price CJ. Detecting subject-specific activations using fuzzy clustering. Neuroimage 2007;36:594-605.

e5.Abdi H, Williams LJ. Principal component analysis. Wiley Interdisciplinary Reviews: Computational Statistics. 2010;2:433-459.

e6.Judd CM, McClelland GH, Ryan CS. Data Analysis: A Model-Comparison

Approach, 2nd ed. New York, NY, US: Routledge/Taylor & Francis; 2008.

e7.Shrout PE, Fleiss JL. Intraclass correlations: uses in assessing rater reliability. Psychol Bull 1989;86:420-428.