1025

View Abstract / View Presentation

Effect of Spatial Inhomogeneity Models on Performance of Machine Learning Based Inversion Algorithms for Brain Magnetic Resonance Elastography

Jonathan M Scott¹, Joshua D Trzasko¹, Armando Manduca¹, Matthew L Senjem¹, Clifford R Jack¹, John Huston III¹, Richard L Ehman¹, and Matthew C Murphy¹
¹Mayo Clinic, Rochester, MN, United States

While the specific spatial inhomogeneity assumption used in training machine-learning MRE inversion algorithms affects stiffness estimates, our in vivo results show only modest impact on repeatability and observed biological effects when relatively small spatial footprints are used.

The piecewise smooth inversion produces the best contrast to noise ratio (CNR) for inclusions in the brain simulating phantom. The piecewise smooth inversion produces the best CNR in 4 of the 6 inclusions and is a close second in a fifth. The only inclusion where the piecewise smooth inversion did not perform well was difficult for all inversions, as no inversion produced a CNR better than 1.

Different inversion material property assumptions produce no difference in repeatability. While stiffness maps in an individual case are appreciably different (bottom row), there is no significant difference in coefficient of variation (CV) between the inversions in any of the 35 gray matter regions of interest. The box plot (line=median, box edges = 25th/75th percentiles, whiskers=extremes, +=outlier) shows the mean of the mean CVs in the 35 regions for the four inversions in 10 subjects.