Synthetic data shows the potential of unsupervised and supervised learning for incorporating spatial information in IVIM fitting

DWI data is spatially homogeneous, yet microstructures are irregular, where neighboring voxels do not always share information. Therefore, we need to utilize neighboring correlations only when they are present. In simulations, we show that by training on synthetic data with all plausible combinations of neighboring correlations, the accuracy of supervised deep learning IVIM model fitting can be improved. Conversely, unsupervised learning did not benefit from incorporating spatial information. In in-vivo data from a glioma patient, supervised training on this synthetic data improved the performance of IVIM fitting by effectively denoising the DWI data while preserving edge-like structures.