Corpus Callosum

Voxel-wise features were extracted in a common template space (Ω_Template, see Figure 2) based on the data of the training set. This common template space was constructed using a procedure that avoids bias toward any of the individual training images (Seghers et al., 2004 ). In this approach, the coordinate transformations from the template space to the subject's image space (V_i: Ω_Template → Ω_Ii) were derived from pairwise image registrations. For computation of V_i, the image of an individual training subject (I_i) was registered to all other training images (I_j) using I_i as the fixed image. This resulted in a set of transformations W_i,j : Ω_Ii→Ω_Ij. By averaging the transformations W_{i, j}, the transformation U_i : Ω_Ii→Ω_Template was calculated:
$U_i(x)=\frac{1}{N}{\displaystyle \sum _{j=1}^N{W}_{i,j}}(x).$
The transformation V_i was calculated as an inversion of U_i: V_i = U_i⁻¹. Note that the identity transformation W_i,i is also included in (5). The pairwise registrations were performed using a similarity (rigid plus isotropic scaling), affine, and nonrigid B-spline transformation model consecutively. The nonrigid B-spline registration used a three-level multi-resolution framework with isotropic control-point spacings of 24, 12, and 6 mm in the three resolutions respectively.
A template image was built using: $\overline{I}(x)=\frac{1}{N}{\displaystyle {\sum }_{i=1}^N{I}_i(V_i(x))},$ with I_i(V_i) representing the deformed individual training images. The test images were not included in the construction of Ω_Template. For the test images, the transformation to template space (V_i) was obtained using the same procedure described above: using pairwise registration of each image with all training images, followed by averaging and inversion. Brain masks and tissue maps were transformed to template space using V_i.
For extraction of the region-wise features, a set of 72 brain ROIs was defined for each subject individually in subject space (Ω_I) using a multi-atlas segmentation procedure (Figure 3). Thirty labeled T1w images containing 83 ROIs each (Hammers et al., 2003 (link); Gousias et al., 2008 (link)) were used as atlas images. The atlas images were registered to the subject's T1w image using a rigid, affine, and nonrigid B-spline transformation model consecutively resulting in transformation S_i,k : Ω_Ii → Ω_Atlask. Registration was performed by maximization of mutual information within dilated brain masks (Smith, 2002 (link)). For initialization, the dilated brain masks were rigidly registered. For nonrigid registration, the same multi-resolution settings were used as in the template space construction. For this step, the subjects' images were corrected for inhomogeneities (Tustison et al., 2010 (link)). Labels were propagated to Ω_Ii using S_i,k and fused using a majority voting algorithm (Heckemann et al., 2006 (link)). The brain stem, corpus callosum, third ventricle, lateral ventricles, cerebellum, and substantia nigra were excluded.

Structural, diffusion, and resting-state functional MRIs were acquired on a Siemens 3T Trio Tim scanner using a 24-channel phased-array head coil. Diffusion tensor imaging and resting-state functional MRI scans were obtained with the same scanner using an identical protocol for all participants during a single visit. The Tracts Constrained by Underlying Anatomy (TRACULA) tool within FreeSurfer 7.1²⁴ was used for diffusion tensor imaging data processing and tractography to estimate the posterior probability of the 18 major white matter tracts. Among the 18 tracts, there were 5 tracts with segments that showed significant differences in both fractional anisotropy (FA) and mean diffusivity (MD) among the ELBW group (with or without PGF): the forceps major of the corpus callosum (Fmajor), right anterior thalamic radiation (RATR), left inferior longitudinal fasciculus, left superior longitudinal fasciculus–parietal bundle (LSLFP), and left superior longitudinal fasciculus–temporal bundle (Figure 1). The CONN toolbox²⁵ was used for the seed-based functional connectivity analysis.^{26 (link)} To select the region of interest as a seed, the multivoxel–multivariate pattern analysis method was used, which showed group differences in 4 regions: the precuneus, the left and right superior lateral occipital cortex (sLOC), and the posterior cingulate cortex (PCC) (eFigure 2 in Supplement 1). A seed-based functional connectivity analysis was performed for whole-brain regions with the selected 4 regions of interest in the multivoxel–multivariate pattern analysis. Functional connectivity strength values were extracted from the brain regions of the preterm infants (uncorrected height threshold of P < .001 and cluster-level false discovery rate–corrected P < .05). The diffusion metrics of the selected tracts and the FCS values were used in the correlational analysis with clinical and neuropsychological measures. Details are provided in the eMethods in Supplement 1.

The comprehensive protocol for data quality management is available in eMethods 1 in Supplement 1. Demographic information, including sex, age, and modified Rankin Scale Score (which measures degree of disability or dependence after a stroke) at hospital admission, was recorded. Radiographic variables describing AVM morphological characteristics, including nidus location, size, diffuseness, venous drainage (drainage patterns, stenosis, and venous aneurysms), feeding arteries (number, dilation, multiple sources, and perforating arteries), associated aneurysm, and hemorrhagic presentation, were collected. Radiological information was determined via digital subtraction angiography and MRI.
The nidus location was regarded as deep if the lesion exclusively involved the brain stem, cerebellum, basal ganglia, thalamus, corpus callosum, or insular lobe. The definition of eloquent regions (ie, sensory, motor, language, or visual cortex; hypothalamus or thalamus; internal capsule; brain stem; cerebellar peduncles [superior, middle, or inferior]; and deep cerebellar nuclei) was based on the Spetzler-Martin Grading Scale.^{5 (link)} The size of AVMs was dichotomized into small and large based on whether the maximum nidal diameter was less than 3 cm or 3 cm or greater. Ventricular system involvement was determined via MRI based on whether the nidal border was adjacent to the cerebral ventricular system. Feeding arteries were considered dilated when their diameter was at least twice that of the same blood vessel segments. Venous aneurysm was defined as the focal aneurysmal dilation of the proximal drainage vein.^{18 (link)} Hemorrhagic presentation was defined as hemorrhage that could be ascribed to AVM rupture before or at admission.