The contribution of lesion location to upper limb deficit after stroke

Background Motor deficit after stroke is related to regional anatomical damage. Objective To examine the influence of lesion location on upper limb motor deficit in chronic patients with stroke. Methods Lesion likelihood maps were created from T1-weighted structural MRI in 33 chronic patients with stroke with either purely subcortical lesions (SC, n=19) or lesions extending to any of the cortical motor areas (CM, n=14). We estimated lesion likelihood maps over the whole brain and applied multivoxel pattern analysis to seek the contribution weight of lesion likelihood to upper limb motor deficit. Among 5 brain regions of interest, the brain region with the greatest contribution to motor deficit was determined for each subgroup. Results The corticospinal tract was most likely to be damaged in both subgroups. However, while damage in the corticospinal tract was the best indicator of motor deficit in the SC patients, motor deficit in the CM patients was best explained by damage in brain areas activated during handgrip. Conclusions Quantification of structural damage can add to models explaining motor outcome after stroke, but assessment of corticospinal tract damage alone is unlikely to be sufficient when considering patients with stroke with a wide range of lesion topography.

We generated a lesion likelihood map for each patient, as follows. The T1-weighted structural images were flipped about the mid-sagittal plane if lesions were in the left hemisphere. All structural images then showed lesions in the right hemisphere and were segmented into seven tissue types, including grey matter, white matter, cerebrospinal fluid, bone, soft tissue, air, and extra tissue, and then normalised using the New Segment routine in SPM8 (http://www.fil.ion.ucl.ac.uk/spm/). The extra tissue type corresponded to the area of infarcted brain and segmentation was repeated twice with an update of the extra tissue type to make sure segmented grey and white matter excluded the lesion. For each patient, the segmented grey and white matter was compared voxel by voxel with 23 healthy controls' segmented grey and white matter, so that the likelihood of damage in each voxel was calculated and expressed as a number between 0 and 1 using the algorithm in the ALI toolbox. 5

Weight map
We then determined how much the likelihood of damage in each voxel contributed to upper limb motor deficit. Here, we employed a machine learning technique using Bayesian inference for regression termed a relevance vector machine. 6 This multi-voxel pattern analysis was performed in each patient subgroup using PRoNTo v1.1 (http://www.mlnl.cs.ucl.ac.uk/pronto/). The multivariate analysis enabled us to assess the association of lesion characteristics with upper limb motor deficit by considering the spatial correlation between all voxels over the grey and white matter, rather than regarding each voxel separately. Leave-one-subject-out cross-validation was performed and the weight with which each voxel contributed to the regression function was summarised as a weight map for each patient subgroup, where a more negative value indicates that a lesion in that voxel is more likely to be associated with greater upper limb motor deficit.

Brain regions of interest
The previous steps provided voxel-wise values of lesion likelihood (one lesion likelihood map per patient) and contribution to motor deficit (one weight map per patient subgroup). Next we were interested in assessing the importance of damage in a number of prespecified brain regions. Five brain regions of interest (ROIs) were defined from healthy controls or the template brain: (1) CST, (2) SM, (3) aSM, (4) GM, and (5) WM. CST was demarcated from healthy controls' DTI data by tracking the fibres from primary and secondary motor cortices including M1, PM, and SMA to ipsilateral lower pons via internal capsule and upper pons as previously described. 1 SM included M1, PM, and SMA as well as primary sensory cortex (S1) as defined in the Harvard-Oxford Atlas. aSM was acquired from healthy controls' fMRI data by determining group level activation during hand grip within SM at an uncorrected p value of 0.001. GM and WM were defined by keeping voxels with probabilities over 0.5 in the grey and white matter probability maps respectively included in SPM8. Note that GM includes basal ganglia and thalamus. More details about defining CST and aSM through analysis of fMRI and DTI data, respectively, are presented below.

Analysis of fMRI data
Preprocessing and statistical analysis of fMRI data were performed using SPM8. The first 6 images were discarded to allow for T1 equilibration effects, so that 114 images were used in preprocessing and statistical analysis. Each subject's 114 images were spatially realigned to the mean image and then unwarped to correct for head motions. The realigned images were normalised into the same coordinate frame as the MNI template brain with transformation parameters derived from segmentation of the subject's high resolution structural image coregistered to the mean functional image. The normalised images were spatially smoothed using a Gaussian filter of 4 mm FWHM. In first level statistical analysis of the preprocessed images, the general linear model included three regressors: (1) hand grip trials modelled as the delta function; (2) an interaction of hand grip with force exerted for it, modelled as the delta function scaled by the peak force; (3) a mean over all images. Each individual subject's activation map for the hand grip task was collected as voxel-wise parameter estimates of the first regressor. In group-level statistical analysis, a group-wise activation map for the hand grip task was acquired by carrying out a one sample t test of all subjects' activation maps. Finally, the group-level activations were thresholded at an uncorrected p value of 0.001 and were confined within sensorimotor cortex including M1, PM, SMA, and S1 to yield a brain region corresponding to healthy controls' activation map.

Analysis of DTI data
Preprocessing, diffusion tensor modelling, and tracking of corticospinal tract were performed using FDT v2.0 included in FSL (http://fsl.fmrib.ox.ac.uk/fsl/). Each subject's 68 images were first realigned to the first image to correct for eddy current-induced distortions and head motions. By Markov Chain Monte Carlo sampling, distributions of voxel-wise principal diffusion directions were inferred. In probabilistic tractography of corticospinal tract, one seed mask, two waypoint masks, one target mask, and one exclusion mask were employed to spatially confine fibres. The seed mask comprised M1, PM, and SMA. The posterior limb of internal capsule (PLIC) and upper and lower pons ipsilateral to the seed mask were manually delineated, of which PLIC and upper pons served as the waypoint masks and lower pons as the target mask. The PLIC mask was placed from the level of anterior commissure to the base of corona radiata, and the upper and lower pons masks were located to only include anterior pons. Corpus callosum and cerebellum were used as the exclusion mask to remove inter-hemispheric and cerebellar trajectories. By repetitively computing 5,000 streamlines starting from every voxel of the seed mask, a distribution of streamline locations from the seed mask to the target mask via the waypoint masks was estimated. In the connectivity distribution, each voxel had a streamline count that passed through the voxel. Each individual subject's corticospinal tract map was acquired as a binary map by thresholding the subject's connectivity distribution at 5% of the maximum voxel value. A group-wise corticospinal tract map was generated by superposing all subjects' corticospinal tract maps and a brain region corresponding to healthy controls' corticospinal tract was determined to only include voxels with overlap counts higher than half of the subject number.

Lesion load and contribution weight of ROIs
Lastly, we wanted to calculate the following: (1) the lesion load of each ROI as the average of voxel-wise values over the ROI in the lesion likelihood map (for each patient) and (2) the contribution weight of each ROI to upper limb motor deficit as the average of voxel-wise values over the ROI in the weight map (for each patient subgroup). The lesion load was compared between each pair of the five ROIs using paired samples t tests for each patient subgroup, and between the two patient subgroup using two samples t tests for each ROI. Statistical significance was determined at a false discovery rate adjusted p value of 0.05. Also, the ROI with the greatest contribution weight to upper limb motor deficit was identified in each patient subgroup. Note that statistical inferences for the weight maps were not feasible, as a weight map was acquired for each patient subgroup, not for each patient.