In this work, we have developed an automated technique to map structural connectivity in the infant brain using diffusion MRI, and used this approach to characterize large-scale connectivity of the cortex in 17 six-month old babies with HIE at birth. The approach is similar to what has been described in adults, but was modified to include more rigorous quality assurance and an anatomically unconstrained approach to parcellation in order to better study the variable anatomy of this group. Interestingly, the derived networks demonstrated properties which were correlated to neuromotor outcome in HIE babies at six months. The choice of the threshold for binarizing connectivity matrices affected the statistical significance of the resulting correlation. Nevertheless, a trend to declining brain network integration and segregation was observed with increasing neuromotor deficits. These results should be interpreted with caution, as the mechanisms for brain network disruption in babies with encephalopathy have not yet been well-characterized and are likely heterogenous in both etiology and outcome 
. A larger number of subjects and a longitudinal study design over years would provide more information on this aspect of the study. The principle result that we intended to achieve is the framework for structural network construction in the pediatric brain – a step towards the large-scale baby connectome that will contribute to our understanding of brain development, as well as developmental abnormalities or lesions that affect brain function.
The important step of data quality assurance preceded the mapping of the structural connectivity. This step included considerations for not only inter-image motion and eddy-current correction, but addressed intra-image diffusion encoding artifacts that were not accounted for by traditional motion correction algorithms. Rejection of diffusion directions in half-Fourier images with artifacts significantly improved fiber tractography. Unless the location of the DC component of the k-space is detected and an adaptive version of the homodyne algorithm is used 
, affected images have to be excluded from diffusion data reconstruction. Furthermore, no algorithm can solve the problem if displacement of the k-space center is outside the sampling range.
No single universal parcellation scheme of the brain exists for the infant brain. A major drawback of using anatomical brain regions, such as the Anatomical Automatic Labeling (AAL) atlas, for studying early brain development was pointed out by Fan et al. 
. Brain networks were built in the same subjects at the ages of 1 month, 1 year, and 2 years based on correlations in regional gray matter volume measures. The authors noted that the AAL atlas might not match very well with function and anatomy of the early development brains. In our study, we proposed two different anatomically unconstrained parcellation schemes. The only anatomic alignment imposed with the proposed gridded parcellation was alignment of the imaging plane with the corpus callosum and interhemispheric fissure, so that the baby brain was evaluated using the same cardinal x
, and z
axes. The equipartition scheme imposed no anatomic constraints. The proposed schemes are straightforward and more suitable for the rapidly changing newborn brain, as they avoid the inherent bias associated with using anatomically-predefined node locations. While the nodes partitioned in the proposed way do not directly correspond to each other across subjects, comparison and assessment are possible using the total resulting network and derived global characteristics. We also expect that, in the future, network-driven co-registration will be an advantageous alternative to atlas-based coregistration, especially in case of challenging age groups and variable anatomy. This will enable unbiased comparison of networks on the local scale, i.e. using single node features. In our study, while both of the parcellation schemes showed the same trend for the small-world metrics, the equipartition demonstrated a stronger correlation with the NMS. The more structured gridded parcellation facilitated visual inspection of the adjacency matrices, in which, e.g. the symmetry of the right and left hemispheres could be easily observed. However, this is only possible in case of a proper alignment of the imaging box with the anatomy. This link to the anatomy, on the other hand, makes the scheme inferior to the equipartition, which is truly automatic and unbiased.
Connections between the cortex and subcortical gray matter structures, such as thalamus, were not analyzed in this work. Including those connections into the analysis would require a relatively precise definition of the inner brain structures manually and/or using templates and, thus, hinder the universal, fully automated approach to studying the developing brain. Results
of mapping of connections between thalamus and cortex in the adult brain using DTI 
and using both DTI and fMRI 
have been reported, but the influence of these structures on overall connectivity is difficult to define.
The proposed framework can be applied to babies of different ages, including premature newborns, and thereby provides a novel tool for unbiased study of structural maturation of the brain. Previously, developmental trajectories could only be studied by measuring anatomy and analyzing separate DTI tracks using tract- or region-of-interest based analysis. We also expect that, by studying brain network topology in newborns, it will become possible to better understand the process of relocation of specific brain functions as a consequence of brain plasticity. The proposed anatomically unconstrained approach to parcellation followed by network-driven analysis of the connectome should facilitate this task.
Recently Hagmann et al. 
applied the principles of MR connectomics to explore the contribution of white matter maturation to the development of connectivity between 2 and 18 years. Among other network refinements, they observed a significant increase in node strength and efficiency along with a decrease in clustering. The betweenness centrality of brain regions remained largely unchanged, with the precuneus, posterior cingulate cortex, superior frontal cortex, and superior parietal cortex remaining the hub regions with the highest centrality ranks. Another very recent longitudinal study by Yap et al. 
explored developmental trends of white matter connectivity in healthy pediatric subjects at ages of 2 weeks, 1 year, and 2 years. The results indicated that the small-world architecture exists at birth with efficiency that increases in later stages of development. The framework developed here specifically aims at facilitating similar studies by ensuring the diffusion data quality and anatomically unbiased parcellation in children under the age of 2.
Our graph theoretical analysis showed small-world properties in six-month old babies. The results are in agreement with previous studies of the adult human brain using EEG, MEG, diffusion MRI, and functional MRI (see 
for a review), as well as with the pediatric studies mentioned in the previous paragraph. While the detection of small-world attributes is considered to be largely independent of the parcellation scheme and spatial resolution 
, the specific network metrics can be affected by both, the network resolution (number and size of nodes), and the angular and spatial resolution of the diffusion acquisition 
. Zalesky et al. 
and Hagmann et al. 
recently showed that the parcellation scale strongly influences the network metrics. We have observed this effect when decreasing the number of nodes from 40 to 20. However, it is also reported that this strong dependence does not suggest that any given parcellation scale is more optimal than another and its choice remains a subject of research.
The patterns of structural connectivity that have been observed in the human brain parallel similar findings of functional connectivity using BOLD fMRI (
for a review of earlier studies). Though interrelated, these two approaches are complementary, and the full description of both structural and functional connectivity is crucial in understanding normal and abnormal maturation of the brain as a whole. Functional connectivity of the newborn brain was studied recently by Fransson et al. 
. It was shown that at the time of birth, the functional brain connectome largely involves brain regions responsible for sensation and action, whereas only weak involvement was found for heteromodal brain areas. The strong candidates for cortical hubs were found in motor, sensory, auditory, and visual primary cortex. Another study in preterm infants 
concluded that all resting state networks, including visual, auditory, somatosensory, motor, default mode, frontoparietal, and executive control networks, are present by term. A recent review by Smyser et al. summarizes exploration of the functional organization of the developing brain 
. However, the importance of the structural network cannot be overemphasized, as functional connections represent a single brain state that unfolds within a milieu of fixed anatomic connections. The combined use of noninvasive structural and functional imaging methods in the same subject would offer the most robust path toward defining the full large-scale connectome. To date, this has been done only for the adult brain, with the structural connectome being the challenging task in the immature brain. The framework developed in this study will facilitate this important step of going from structure to function, which is essential for understanding how cognitive processes emerge from their morphological substrates 
In the present study, we used diffusion tensor MRI in combination with deterministic tractography to track white matter pathways. Though fast and straightforward, deterministic tractography produces reliable results only in brain areas where anisotropy is high 
. As fibers approach the cortex, diffusion anisotropy diminishes, and calculated principal diffusion directions become increasingly uncertain as a result 
. This has limited attempts to trace pathways directly from deep gray matter, which typically has low anisotropy. To reduce the effect of this limitation, we restrained connectivity mapping to the white matter by choosing the nodes on the subcortical surface 2–4 mm below the cortex. Probabilistic tractography could be used instead to improve fiber tracking. Yo et al. showed that probabilistic approaches show on average more connected regions but lower connectivity values than deterministic methods 
. High angular resolution diffusion models may also reveal connections between more brain areas than the simple tensor model, by resolving crossing fibers. These differences should be taken into account when comparing results obtained with different frameworks for assembling the connectome.