Functional connectivity, defined as the coordination of activity across anatomically distinct regions of the brain (Aertsen et al., 1989; Friston et al., 1993), is based on assessment of neuronal activity, either directly, through electrophysiological signals, or indirectly, via hemodynamics or metabolites fluctuations.

Functional MRI (fMRI), arguably the most commonly used technique to assess brain activation, relies on the measurement of blood-oxygen-level-dependant (BOLD) signals. It reflects changes of blood volume and deoxyhemoglobin (HbR) concentrations (Buxton, 2013). fMRI was first described by Ogawa in 1990 (Ogawa and Lee, 1990; Ogawa et al., 1990) and has since revolutionized human neuroscience by allowing the identification of brain areas responsible for diverse cognitive task or processing various stimuli (Belliveau et al., 1991; Raichle, 2009). Resting-state fMRI (rs-fMRI) introduced a few years later (Biswal et al., 1995) corresponds to the measurement of low frequency fluctuations (0.009 < f < 0.1 Hz) in the BOLD signal in the absence of a task or stimulus. It has also greatly contributed to the neuroscience field by revealing the existence of multiple spatially distributed large-scale networks in the brain, the resting-state networks (RSN). Lewis and al. have shown that a possible function of RSN could be related to the consolidation of previous experience (Lewis et al., 2009). According to another hypothesis, the networks functioning during active processing are maintained during rest (and therefore become the RSN) in order to allow rapid activation when needed for active processing (Smith et al., 2009). However, the true function of the RSN remains unknown.

The application of fMRI or rs-fMRI in animal models has been scarce and mostly limited to rats and monkeys, because smaller models such as mice or neonatal rats require a very high intensity magnetic field to obtain sufficient signal-to-noise ratio and spatial resolution to assess their small brains (Benveniste and Blackband, 2002; Jonckers et al., 2011). Given the tremendous number of already well-characterized models that can't be readily studied because of this limitation, interest for the other functional neuroimaging techniques, alone or in combination with fMRI, has increased over the past years (Cang et al., 2005; He and Liu, 2008; Ye et al., 2016). Optical imaging of intrinsic signals (OIS) was shown to be a potent alternative to fc-MRI, also by measuring changes in blood oxygenation (Grinvald et al., 1986; Frostig et al., 1990; Ts'o et al., 1990). In this technique, two wavelengths based on the absorption spectra of oxyhemoglobin (HbO₂) and deoxyhemoglobin (HbR), are shone upon exposed nervous tissue and fluctuations in reflected light intensity are recorded. Those fluctuations represent changes in blood oxygenation/volume and, therefore, can be used to infer changes in neuronal activity. Advantages of OIS over (rs-)fMRI include a much higher spatial resolution (in the order of tens of micrometers for OIS vs. 200–400 μm for fMRI) and temporal resolution (30 Hz for OIS vs. 0.5 s for high-resolution fMRI sequences; Pouratian and Toga, 2002; Jonckers et al., 2015; Pan et al., 2015), providing an interesting tool to assess functional connectivity in small animals models. Unlike fMRI, OIS is not prone to artifacts related to mechanical vibrations and spurious responses arisen from loud acoustic stimuli. Moreover, OIS also uses non-ionizing radiation and is much more cost-effective than fMRI. However, one major limitation of OIS is its inability to assess subcortical structures (Hillman, 2007). The resting state data, acquired with this technique, is hence 2 dimensional. Nevertheless, most of the RSN have some, if not all, of their brain regions in the cortex (Rosazza and Minati, 2011), making OIS a suitable method to study RSNs in small models, like rodents (Li et al., 2012; Guevara et al., 2013c).

The study of alterations in RSN via rs-fMRI has improved our understanding of many neurological and psychiatric diseases such as epilepsy (Wang Z. et al., 2011; Mankinen et al., 2012; Tracy and Doucet, 2015), depression (Mulders et al., 2015), autism (Farrant and Uddin, 2016; Mevel and Fransson, 2016), schizophrenia (Sheffield and Barch, 2016), and dementia (Greicius et al., 2004; Gili et al., 2011; Peraza et al., 2015). The impact of prematurity on neurocognitive development has also been, more recently, studied through fMRI and rs-fMRI. Studies have identified resting state network maturation in the growing brain with evidence of networks as early as 30 weeks of gestational age with a fast maturation leading to adult-like network at term equivalent age (Doria et al., 2010; Smyser et al., 2010; Lee et al., 2013; van den Heuvel et al., 2015; He and Parikh, 2016). The consequences of prematurity itself was revealed by showing a decreased functional connectivity of RSN in (very preterm) VPT-born patients (Smyser et al., 2010, 2013; Ye et al., 2016), best seen using correlation and covariance matrix analyses demonstrated by Smyser et al. (2016b). In other words, the topography of RSN in VPT-born children seems preserved, but quantitative parameters, such as the synchronicity between networks (White et al., 2014; Ye et al., 2016), the amplitude of BOLD signal fluctuations (Smyser et al., 2013, 2016b) or the number of voxels in a RSN (Smyser et al., 2010), appeared to be decreased with a more pronounced reduction in higher order RSN (Default mode, executive control, and frontoparietal networks; Smyser et al., 2016b). Most studies reported that primary RSN (somatosensory, motor, and visual) are almost fully formed (i.e., adult-like) before week 30 of PMA while higher order networks mature during the last trimester (Doria et al., 2010).

The exploration of RSN in VPT children is still in its very first steps. Studies have investigated connectivity in the first 4 years of life of children born preterm (Lee et al., 2013) and in adults and older children born preterm (White et al., 2014; Ye et al., 2016). In spite of the useful information obtained through cohort studies, preclinical models are crucial to test neuroprotective treatment. However, as stated earlier, the data on rodents or other small models remain scarce because of the intrinsic limitations of fMRI. In this study, we characterized using rs-OIS, resting state networks in a preclinical model of diffuse white matter injury that consists in the intra-cerebral injection of lipopolysaccharide (LPS) (Cai et al., 2003; Pang et al., 2003; Lodygensky et al., 2010, 2014; Guevara et al., 2013a). LPS is known to mimic every hallmark of inflammatory white matter injury, in the immature brain such as seen in preterm infants following necrotizing enterocolitis or born in the context of severe chorioamnionitis e.g., hippocampal volume impairment (Wang et al., 2013), ventricle dilation, apoptosis, pre-olygodendrocite cell necrosis, white matter rarefaction, hypomyelination, microglial reaction (Cai et al., 2003; Pang et al., 2003) as well as neurobehavioral deficits (Fan et al., 2005).

The optical intrinsic signal obtained from young rat allowed the characterization of resting state homeostasis in normal and inflammatory conditions. Three complementary approaches were used: (i) Seed-based functional connectivity analysis of rs-OIS; (ii) Support vector machine; (iii) Artificial Neural Network.

Optical intrinsic signal represents a major advantage over fcMRI due to a very high temporal and spatial resolution. It is straightforward to setup compared to MRI, with neither associated mechanical vibrations nor loud sounds, thus avoiding vibration and acoustic noise-related artifacts; while providing better access to the animal, thus facilitating its manipulation. It is obviously restricted to 2D mapping and not a technique of choice for deep brain structures. This is an important limitation when performing regression on rs-OIS data; there is the possibility that global signal regression leads to spurious negative correlations (Murphy et al., 2009; Saad et al., 2012), although some works suggest a physiological basis for anticorrelated networks (Fox et al., 2009). Unfortunately, a viable alternate option is not provided by the OIS technique. Further work is needed on this issue.

As shown in adult mice (Guevara et al., 2013c) we found also that HbR contrast appears to be more sensitive to uncover resting state networks in the young rat brain (Figure 1). Interestingly HbR has been shown to closely match the BOLD contrast in fMRI experiments (Huppert et al., 2006).

When compared to adult mice (Guevara et al., 2013c), rs-OIS in young rats detected similar functional correlation maps, across different seeds, showing high degree of inter-hemispheric symmetry in control animals.

In the developing rat brain, an inflammatory brain injury was found to cause significant long standing disruption of resting state activity using rs-OIS with a reduction of connectivity in several networks particularly on motor and cingulate networks. The clear impact of inflammation on these networks is not unexpected considering the site of LPS injection. Nevertheless, the identification of persisting deficits is remarkable as these results are in agreement with preterm infants with substantial white matter injury, with a significant reduction on motor networks and in the vicinity of clear injury (Smyser et al., 2013).

When studying averaged seed-based correlation maps, inflammation was shown to cause a significant decrease of spatial extent across cortical regions, that was highly significant on HbR maps (Figures 2–4). This difference in HbR with respect to HbO₂ may be explained by the higher contribution of HbR to absorption at 590 and 630 nm wavelengths. As mentioned in the introduction similar reduction of correlated voxels where found in preterm infants when compared to term infants imaged at term equivalent age (Smyser et al., 2010). Translating this approach on pre-clinical fcMRI in young developing rat or mice brains might prove difficult as partial volume effect might dampen differences between groups especially in small correlated regions such as in the cingulate cortex, where fc-OIS was able to find a difference of ~0.6 mm². This small, albeit significant difference might be challenging to unravel using fcMRI in rodents, where voxel size is usually between 0.5 mm³ (Liang et al., 2012) and 1 mm³ (Harris et al., 2015).

The use of a SVM allowed us despite the limited sample size, to efficiently classify injured from intact developing brains with accuracy above 90%. In contrast to seed based analysis it provides a categorical result with no indication regarding the type of changes. The advantages and limitations still need to be properly addressed especially if used to determine therapeutic efficiency. Nevertheless, it is a first step in detecting injury through optical imaging, since SVM was shown to identify injury with high accuracy, using rs-OIS data. Similarly, Vergara et al. have used SVM with a leave-one-out cross validation to classify mild traumatic brain injury based on fcMRI with an accuracy of 84.1 % that was superior to DTI analysis (Vergara et al., 2017). This approach was also successfully applied in a cohort of preterm infants also using fcMRI with an 84% accuracy and a specificity of 78% (Smyser et al., 2016a). The range of accuracy probably depends on the severity of injury. It is noteworthy that this approach remains extremely efficient in the setting of mild traumatic brain injury and in a cohort of healthy preterm infants at term with no overt brain injury. The increased temporal resolution of rs-OIS over fcMRI may increase the quality of resting state assessment. It is possible that in young rodent, rs-OIS might reveal itself superior in comparative studies with fcMRI.

The use of an ANN allowed us to assess the degree of injury with a highly significant correlation with the degree of ventricular dilation. Machine learning approaches have proven useful in neonatal populations, e.g., using SVM to predict gestational age (Smyser et al., 2016a), ANN to assess mortality risk (Zernikow et al., 1998) or heuristic methods to detect seizures (Ansari et al., 2016). One limitation of ANN, and machine learning in general, is the fact that the only observed states are the inputs and the outputs, making interpretation of their inner functions exceedingly difficult (Castelvecchi, 2016). Considering the small number of animals, rs-OIS appears to be a highly reliable tool, sensitive enough to pick up differences in lesion load.

An analysis of resting state networks in the developing rodent brain together with the impact of inflammatory white matter injury was presented using fc measurements and machine learning techniques. Our approach was based on creating group-level fc statistical maps where a significant decrease in connectivity was observed in the case of the injured group; furthermore, seed-based fc measures were analyzed through SVM to identify connectivity patterns differences between NaCl and LPS groups. Those same fc measures were fed to an ANN that allowed the prediction of fractional lesion volume. Different fc approaches (ICA, graph theory, clustering, etc.) and alternatives to the chosen machine learning algorithms need further investigation on the basis of a larger sample. Another drawback of this study is the assumption of stationarity in fc, thus requiring the consideration of dynamic models in future studies (Hansen et al., 2015). Nevertheless, the translation of both machine-learning approaches, to a bedside optical monitoring system, such as fNIRS, could be helpful in providing early diagnosis in preterm infants especially when evaluating small cohorts.

EG, FL, and GL conceived and designed the experiments; WP, CT, LA, IL, and GL performed the experiments; EG, CT, LA, IL, and GL analyzed the data; FL and GL contributed materials and analysis tools; EG, CT, FL, and GL wrote the paper.