Early brain growth is characterized by rapid and complex structural and functional developments that are vulnerable to various genetic and environmental factors. The influence of early brain development and disorders on brain health later in life has received growing interest (Volpe, 2009, Bhat et al., 2014, Kwon and Kim, 2017, O’Donnell and Meaney, 2017, Bilder et al., 2019). Magnetic Resonance Imaging (MRI) is a non-invasive method for assessing these developments in vivo. Diffusion MRI (dMRI), specifically, offers a means to assess the microstructure of the white matter using the diffusion of water molecules as a proxy measure (Stejskal and Tanner, 1965, Le Bihan et al., 1986). However, the application of dMRI to study the developing brain has been limited due to motion, limited scan time, and low signal-to-noise ratio (SNR) (Dubois et al., 2014, Christiaens et al., 2019, Jakab et al., 2015b). Despite these limitations, prior works have shown the potential of dMRI to probe early brain development. For instance, several studies (Wilson et al., 2021, Xu et al., 2022, Calixto et al., 2023) have used spatiotemporal changes in Fractional Anisotropy (FA), Mean Diffusivity (MD) and different cortical morphology indices to characterize the normal brain development. The recent availability of large high-quality datasets, such as those collected under the developing Human Connectome Project (dHCP) (Hutter et al., 2018, Tournier et al., 2020), presents a unique opportunity to enhance our understanding of the developing brain. These datasets include dense multi-shell measurements. As such, derived dMRI quantities can be considered as reference values to which derived metrics from more constrained clinical datasets, which usually do not exceed 15 diffusion measurements with a single low b-value (500–750s/mm2), can be compared.

The prevailing way of extracting diffusion properties from the diffusion signal involves a model, typically a diffusion tensor imaging (DTI) model (Basser et al., 1994). More complex models such as the multi-shell multi-tissue constrained spherical deconvolution (MSMT-CSD) (Tournier et al., 2004, Jeurissen et al., 2014) aim to reconstruct Fiber Orientation Distribution Functions (FODs) that allow depiction of more intricate white matter configurations such as fiber crossings. These models require densely sampled multi-shell data. The output of all of these models can be studied directly, i.e. by computing metrics such as FA or MD from the diffusion tensor or the apparent fiber density (Raffelt et al., 2012) from the FOD. Alternatively, they can be further processed globally to reconstruct the fiber tracts (Jbabdi and Johansen-Berg, 2011, Jeurissen et al., 2019) that are responsible for transmitting action potentials between different regions of the brain.

In general, mapping the acquired diffusion signal to an interpretable and informative diffusion metric requires a prior model. Conventional estimation methods do not provide feedback regarding which measurements are more informative for estimating the given model. Differently, deep neural networks can treat the problem as a single learnable task that can be optimized via back-propagation by directly learning a mapping between the diffusion signal and the target diffusion quantity. Hence, they avoid the sub-optimal model fitting step that can be sensitive to noise. Golkov et al. (2016) proposed the first deep learning (DL) model that directly estimated diffusion kurtosis (Lu et al., 2006) and neurite orientation dispersion and density measures (Zhang et al., 2012) from a small number of diffusion measurements in adult brains. They showed a drastic decrease in scanning time with limited loss in accuracy.

Differently, approaches optimizing the raw diffusion signal instead of directly learning end-goal quantities (i.e. tensors, scalar maps, FODs, tracts) with deep learning have been proposed. Such approaches can operate on k-space, where for instance Mani et al., 2020, Mani et al., 2022 aimed to accelerate acquisitions and implicitly increase the angular and spatial resolution by operating in the k-space and q-space domains. Other approaches operating in the image space such as (Yin et al., 2019, Lyon et al., 2022) have aimed at estimating new gradient directions from acquired ones.

However, if the goal is to compute a specific micro-structure index, a direct optimization approach may be more effective. Since the pioneering study of Golkov et al. (2016), several other works have explored DL methods in adult brains as to directly estimate diffusion scalars. For instance, with superDTI (Li et al., 2021), accurate predictions of tensor maps using a neural network were achieved using only six diffusion measurements. This model was robust to various noise levels and could depict lesions present in the dataset. Koppers and Merhof (2016) employed a 2D convolutional neural network (CNN) in a classification approach to predict the orientation of fibers, while Lin et al. (2019) utilized a 3D CNN to predict FODs based on a small neighborhood of the diffusion signal. Nath et al. (2019) employed a dense residual network to predict FODs obtained from ex-vivo confocal microscopy images of animal histology sections. This approach is limited due to the scarcity of ex-vivo histological training data. Karimi et al. (2021b) used a multilayer perceptron (MLP) to predict FODs. To leverage correlations between neighboring voxels, Hosseini et al. (2022) used a two-stage Transformer-CNN to map 200 measurements to 60 measurements, followed by predicting FODs. Nonetheless, acquiring such a large number of measurements is difficult and frequently infeasible for noncooperative cohorts, such as neonates or fetuses.

The challenge of acquiring useful dMRI data from newborns and fetuses who tend to unpredictably move during long and loud dMRI acquisitions is exacerbated by the low signal available from the small size of the immature and developing brain compartments, the low resolution of dMRI, and the rapid and large changes that occur to the brain microstructure across gestation and early after birth.

These complex maturation processes that are unfolding during gestation include the development of major fiber bundles, namely limbic and projection fibers, during the first trimester (Khan et al., 2019). For instance, the internal capsule experiences intricate microstructure alterations as a result of the intertwining of multiple fiber pathways that initiate development during different periods of gestation. In the second trimester, association fibers start developing and become evident in the third trimester. Specifically, the superior longitudinal fasciculus exhibits accelerated growth during this phase and continues to undergo substantial development even beyond the time of birth (Huang and Vasung, 2014). The radial coherence within the telencephalic wall gradually diminishes with gestational weeks (GW). Furthermore, the regional radial coherence within the deep subplate zone starts to vanish around 26 GW. This regional loss of radial coherence aligns temporally with the previously reported emergence of long-association cortico-cortical tracts (Vasung et al., 2017).

These dynamic changes and the other aforementioned problems of this sensitive population, added to B0 and B1 inhomogeneities, pose additional challenges for learning-based methods for FOD estimation, such as those reviewed above. Therefore, in this study, we build upon our preliminary work in Kebiri et al. (2023), to extensively investigate the use and generalizability of deep learning to estimate the microstructure of the developing brain in different cohorts of newborns and fetuses from 24 weeks of gestation to 48 post-menstrual weeks. To the best of our knowledge, these learning-based FOD estimation methods have not yet been critically evaluated for fetal populations and in non-research protocols of newborns. In this study, we demonstrate that a deep convolutional neural network with a large field of view (FOV) can accurately estimate FODs using only 6–12 diffusion-weighted measurements. Firstly, we show, on N=465 subjects from the dHCP dataset, that our deep learning approach outperforms current state-of-the-art deep learning methods using the same number of measurements, and can achieve a level of accuracy that is comparable to the accuracy of the state-of-the-art standard methods while reducing the required number of measurements by a factor of ∼21–43. Secondly, we present evidence of a low agreement among standard FOD estimation methods for these age groups and show that by increasing the number of input measurements in our method, we approach this upper bound agreement. Thirdly, we show the generalizability of deep learning methods on two out-of-domain clinical datasets of 26 in vivo fetuses and neonates that were scanned with different scanners and acquisition protocols. Finally, we assess for the first time, the deep learning-generated fetal FODs with post-mortem histological data of corresponding gestational weeks.