Conventional MR images and pseudo-CT’s (pCT’s) generated using state-of-the-art machine learning techniques poorly characterize bone anatomies, preventing applicability for orthopedic applications. We hypothesize that smart use of several specific MR contrasts will expose the information needed for diagnostic quality bone visualization. We designed a patch-based convolutional neural network taking groups of different MR contrasts -which were obtained from a single multi-gradient sequence- as inputs . It generated competitive pCT scans, capturing local anatomical variances present in the dataset. We show that Dixon reconstructed inputs appear to generate better soft-tissue visualization, while complex-valued data show promising results in bone reconstruction.
This abstract and the presentation materials are available to members only; a login is required.