19 healthy volunteers underwent 3 successive MRI
sessions (total timespan 2-3h), in which we acquired 3D mDIXON images in exhale
breathhold with an empty (3h of fasting), half-full (half a meal) and full
stomach (remaining half meal). Images of each volunteer were rigidly registered
on bony anatomy. Stomachs were delineated and triangulated meshes were created
(Fig 1). For both strategies (Fig 2), the non-rigid iterative closest point
algorithm was used as deformable registration between meshes.
For the personalized strategy, the empty stomach
was mapped to the half-full stomach for each volunteer. This vector field was
extrapolated to predict a full stomach of equal volume to the true full
stomach.
For the population-based strategy, deformation
vectors were acquired between half-full and full stomach for each volunteer. Using
rigid and then deformable registration, we acquired point-to-point
correspondence between every half-full stomach and a reference stomach. Using
this correspondence, deformation vectors were transferred to the corresponding points
on the reference stomach, resulting in 19 deformation vectors to a full stomach
for every point on the reference stomach. For each volunteer, a population-based
full stomach was created by applying scaled average deformation vectors of the
18 other volunteers to the half-full stomach (using the point-to-point
correspondence) to predict a full stomach of equal volume to the true full
stomach.
We evaluated strategy performance by comparing the
following parameters between predicted and true full stomach (paired tests):
Hausdorff distance (dH), 75% of nearest neighbor distances from predicted
to true stomach (d75%), Dice coefficient, missed volume (i.e., volume
of true stomach outside predicted stomach + 10-mm uniform margin).