fix(resampling): Use nearest mode for extrapolating data outside image boundaries (#3453)

This PR addresses an interpolation effect when the brain signal goes
right up to the edge of the image.

We have used the
[scipy.ndimage.map_coordinates](https://docs.scipy.org/doc/scipy/reference/generated/scipy.ndimage.map_coordinates.html)
default interpolation mode of `"constant"`, which extends the input
image with zeros when necessary to perform interpolation (within two
voxels of the edge, for the cubic interpolation we do). This will
frequently result in zeroed-out portions of edge voxels.

This generally works well when the image field of view is large enough
to fully contain the brain, but can cause problems for limited
field-of-view scans or images where the brain tissue can be found in the
edge voxels.

By switching to `nearest`, the edge voxels are extended out. For images
with plenty of space around the head, this should have very little
impact. When brain tissue extends to the image boundary, this should
avoid inducing a signal drop-off.

Closes #3452.

Author: effigies

fmriprep/interfaces/resampling.py

@@ -54,8 +54,15 @@ class ResampleSeriesInputSpec(TraitedSpec):
     num_threads = traits.Int(1, usedefault=True, desc='Number of threads to use for resampling')
     output_data_type = traits.Str('float32', usedefault=True, desc='Data type of output image')
     order = traits.Int(3, usedefault=True, desc='Order of interpolation (0=nearest, 3=cubic)')
-    mode = traits.Str(
+    mode = traits.Enum(
+        'nearest',
         'constant',
+        'mirror',
+        'reflect',
+        'wrap',
+        'grid-constant',
+        'grid-mirror',
+        'grid-wrap',
         usedefault=True,
         desc='How data is extended beyond its boundaries. '
         'See scipy.ndimage.map_coordinates for more details.',