Unsupervised machine learning for free water-eliminated diffusion tensor imaging of the fetal brain.

Accurate characterization of the fetal brain tissue microstructure with diffusion MRI remains challenging due to the presence of significant free water contribution and low data quality inherent in fetal imaging. To address these challenges, in this work we have developed a machine learning method for Free Water Eliminated Diffusion Tensor Imaging (FWE-DTI) that explicitly distinguishes between free and tissue-bound water diffusion. We leverage research-quality neonatal and fetal diffusion MRI data in a stage-wise learning approach that uses supervised and unsupervised training. The effectiveness of our method was quantitatively and qualitatively assessed using in-utero data. In addition to research-quality data used for training and testing, we included lower-quality external data for independent validation. Compared with conventional estimation techniques, our method showed significant reductions in estimation error for fractional anisotropy, mean diffusivity, and free-water fraction, and more stable parameter estimates under realistic measurement sub-sampling. These results highlight the accuracy and robustness of the proposed method in probing tissue characteristics, even with limited data quality, thereby enabling a more reliable assessment of the developmental changes in the fetal brain. Qualitatively, our method yielded visually better diffusion tensors and improved tractography. Overall, our experiments show that the new method is useful for studying subtle variations in normal and abnormal fetal brain development.