Multimodal image analysis of clinical influences on preterm brain development

Abstract

Objective: Premature birth is associated with numerous complex abnormalities of white and gray matter and a high incidence of long-term neurocognitive impairment. An integrated understanding of these abnormalities and their association with clinical events is lacking. The aim of this study was to identify specific patterns of abnormal cerebral development and their antenatal and postnatal antecedents.

Methods: In a prospective cohort of 449 infants (226 male), we performed a multivariate and data-driven analysis combining multiple imaging modalities. Using canonical correlation analysis, we sought separable multimodal imaging markers associated with specific clinical and environmental factors and correlated to neurodevelopmental outcome at 2 years.

Results: We found five independent patterns of neuroanatomical variation that related to clinical factors including age, prematurity, sex, intrauterine complications, and postnatal adversity. We also confirmed the association between imaging markers of neuroanatomical abnormality and poor cognitive and motor outcomes at 2 years.

Interpretation: This data-driven approach defined novel and clinically relevant imaging markers of cerebral maldevelopment, which offer new insights into the nature of preterm brain injury. Ann Neurol 2017;82:233-246.

Figure 1

Analysis pipeline. Multimodal imaging data sets were decomposed into a set of linked independent components (A). Each component represents a spatially independent pattern of variation linked across modalities by a shared subject course, or component weight. Subject‐specific weights for all components were concatenated before CCA. The canonical correlation analysis is illustrated in (B). Multivariate associations between clinical and imaging data are sought by calculating model weights, a and b, that maximize the correlation between the clinical and imaging variates, U and V. Subsequent canonical pairs are sought under that constraint that they are orthogonal. Clinical and imaging loadings are calculated by correlating the canonical variates with the original clinical and imaging data sets. CCA = canonical correlation analysis; FA = fractional anisotropy; ICA = independent component analysis; MD = mean diffusivity.

Figure 2

Imaging patterns correlating with age at scan and low birth weight. Image loadings associated with the first two canonical pairs (A: pair 1; B: pair 2). Maps are shown thresholded at p < 0.01 (corrected for multiple comparisons), the colour bar indicates the t‐statistic. Warm colours show regions positively correlated with the canonical variate, and vice versa. The correlation between the canonical imaging and clinical variates for each pair are shown in the scatterplot in red. Inset: scatterplot (blue) of the correlation between the imaging variate score and the clinical factor with the largest clinical variate loading. Unthresholded statistical maps are available to view online at http://neurovault.org/collections/2178. a.u. = arbitrary units; FA = fractional anisotropy; MD = mean diffusivity.

Figure 3

Imaging patterns correlating with male sex, intrauterine growth restriction, and postnatal illness. Imaging markers for canonical pairs 3 (A), 4 (B) and 5 (C). Images are displayed as in Figure 2. Modalities are only displayed if any voxels showed a significant association with the corresponding imaging variate (p < 0.01). Unthresholded statistical maps are available to view online at http://neurovault.org/collections/2178. a.u. = arbitrary units; FA = fractional anisotropy; MD = mean diffusivity.

Figure 4

Comparing imaging markers with and without the addition of intracranial volume to the CCA analysis. The addition of ICV to the model results in an additional canonical pair (not shown) detailing the relationship between each modality and total brain volume. As subsequent canonical pairs are defined independent of this relationship, the right column shows each imaging marker corrected for ICV. CCA = canonical correlation analysis; FA = fractional anisotropy; ICV = intracranial volume; MD = mean diffusivity.