Diffusion magnetic resonance imaging has emerged as the gold standard tool for studying the brain white matter both in vivo and non-invasively, offering valuable insights into underlying tissue microstructure and brain connectivity. However, applying this technique to investigate the human developing brain, such as in fetuses and newborns, poses unique challenges. In this sensitive population, the scanning time is limited for unpredictable motion risk minimization and for ethical reasons. Additionally, images have a low signal-to-noise ratio and a low spatial resolution. Moreover, the developing brain undergoes rapidly changing microstructural properties during the last months of pregnancy and early months of birth. The application of current diffusion magnetic resonance imaging methods to developing brains is severely constrained by all these aspects, necessitating the development of tailored approaches. This thesis tackles this specific problem by proposing two deep learning based methods that leverage high quality research datasets to improve constrained clinical acquisitions. First, we have developed a method to enhance the through-plane resolution using a deep autoencoder. We show its performance over conventional image interpolation methods of the raw signal and in estimated diffusion tensor scalar maps. Second, we designed a model to predict accurate orientation distribution functions from a low number of diffusion measurements that are typically available in clinical settings. We extensively demonstrate its performance on newborn subjects compared to state-of-the-art methods (such as constrained spherical deconvolution) that need significantly more diffusion directions. We additionally show the out-of-domain generalizability of the method on clinical cohorts of newborns and fetuses. Finally, aiming at deriving optimal schemes for fetal sequences, we have conducted a quantitative validation study on a phantom with crossing-fibers, to quantify the time trade-off that is imposed by the clinical constraints, between the number of gradient directions and the number of acquired volumes.
Overall, we believe that the aforementioned methods that harness the capabilities of deep neural networks to extract transferable knowledge from large datasets, possess the potential to offer significant insights into the complex mechanisms underlying the early development of the human brain.