The relatively low resolution of conventional in-vivo human dMRI prevents the reconstruction of small axonal bundles (diameter < 3mm). Here we show that diencephalic connections, which are inaccessible at macroscopic resolution (1.5mm), can be reconstructed with unprecedented accuracy at mesoscopic resolution (760μm). We investigate the minimum number of directions needed to achieve this with a state-of-the-art, multi-slab dMRI sequence. We find that a long, multi-session acquisition is necessary. We aim to develop a protocol for manual annotation of these bundles in a small number of high-quality datasets, and thus produce a training set for automated reconstruction in lower-quality dMRI data.