Magnetic resonance imaging (MRI) can be sensitive to tissue microstructural features and infer parameterized features by performing a voxel-wise fit of the signal to a biophysical model. However, biophysical models rely on simplified representations of brain tissue. Machine learning (ML) techniques may serve as a data-driven approach to optimize for microstructural feature extraction. Unfortunately, training an ML model for these applications requires a large database of paired specimen MRI and histology datasets, which is costly, cumbersome, and challenging to acquire. In this work, we present a novel approach allowing a reliable estimation of brain tissue microstructure using MRI as inputs, with a minimal amount of paired MRI-histology data. Our method involves pretraining a conditional normalizing flow model to predict the distribution of microstructural features. The model is trained on synthetic MRI data generated from unpaired histology and MRI physics, reducing the data requirement in future steps. The synthetic MRI generation data combines segmentation of a publicly available EM slice, feature extraction and MRI simulators. Subsequently, the model is fine-tuned using experimental MRI/Electron Microscopy (EM) data of nine excised mouse brains through transfer learning. This approach enables the prediction of non-parameterized joint distributions of g-ratio and axon diameters for a given voxel based on MRI input. Results show a close agreement between the distributions predicted by the network and the EM ground-truth histograms (mean Jensen-Shannon Distances of 0.24 and 0.23 on the test set, for axon diameter and g-ratios respectively, compared to distances of 0.18 and 0.18 of a direct fitting of a Gamma distribution to the ground truth). The approach also shows up to 4% decreased mean percent errors of the distributions compared to biophysical model fitting and increased prediction capabilities that are consistent with electron microscopy validation and previous biological studies. For example, g-ratio values predicted along the corpus callosum anterior-posterior axis show a significant difference for mice after myelin remodeling seizures are well established (p < 0.001) but not before seizure onset (p = 0.562). The results suggest that pretraining on synthetic MRI and then using transfer learning is an effective approach for addressing the lack of paired MRI/histology data when training ML models for microstructure prediction. This approach is a step toward developing a versatile and widely used foundation model for predicting microstructural features using MRI.