AbstractCompressed sensing (CS) has been used to enhance the feasibility of diffusion spectrum imaging (DSI) by reducing the required acquisition time. CS applied to DSI (CS-DSI) attempts to reconstruct diffusion probability density functions (PDFs) from significantly undersampled q-space data. Dictionary-based CS-DSI using L2-regularized algorithms is an intriguing approach that has demonstrated high fidelity reconstructions, fast computation times and inter-subject generalizability when tested on in vivo data. CS-DSI reconstruction fidelity is typically evaluated using the fully sampled data as ground truth. However, it is difficult to gauge how great an error with respect to the fully sampled PDF we can tolerate, without knowing whether that error also translates to substantial loss of accuracy with respect to the true fiber orientations. Here, we obtain direct measurements of axonal orientations in ex vivo human brain tissue at microscopic resolution with polarization-sensitive optical coherence tomography (PSOCT). We employ dictionary-based CS reconstruction methods to DSI data from the same samples, acquired at high max b-value (40000 s/mm2) and with high spatial resolution. We compare the diffusion orientation estimates from both CS and fully sampled DSI to the ground-truth orientations from PSOCT. This allows us to investigate the conditions under which CS reconstruction preserves the accuracy of diffusion orientation estimates with respect to PSOCT. We find that, for a CS acceleration factor of R=3, CS-DSI preserves the accuracy of the fully sampled DSI data. That acceleration is sufficient to make the acquisition time of DSI comparable to that of state-of-the-art single- or multi-shell acquisitions. We also show that, as the acceleration factor increases further, different CS reconstruction methods degrade in different ways. Finally, we find that the signal-to-noise (SNR) of the training data used to construct the dictionary can have an impact on the accuracy of the CS-DSI, but that there is substantial robustness to loss of SNR in the test data.