Merge pull request #44 from eilidhmacnicol/enh/brain_extraction_optim…

…ization

ENH: improved brain extraction parameters

Dockerfile

@@ -77,9 +77,9 @@ ENV PATH=/usr/local/miniconda/bin:$PATH \
 RUN conda install -y python=3.7.1 \
                      mkl=2018.0.3 \
                      mkl-service \
-                     numpy=1.15.4 \
+                     numpy>=1.16.5 \
                      scipy=1.1.0 \
-                     scikit-learn=0.19.1 \
+                     scikit-learn>=0.20 \
                      matplotlib=2.2.2 \
                      pandas=0.23.4 \
                      libxml2=2.9.8 \

nirodents/cli/brainextraction.py

@@ -122,16 +122,6 @@ def get_parser():
         default=(0, 4, 4),
         help="Specify ants-ai Search Grid parameter",
     )
-
-    parser.add_argument(
-        "--slice-direction",
-        dest="slice_direction",
-        action="store",
-        type=int,
-        default=1,
-        choices=[0, 1, 2],
-        help="Specify slice direction: 0 = Sagittal (R-L), 1 = Coronal (A-P), 2 = Axial (S-I)",
-    )
     return parser
 
 
@@ -148,7 +138,6 @@ def main():
         arc=opts.antsai_arcfrac,
         step=opts.antsai_step,
         grid=tuple(opts.antsai_grid),
-        slice_direction=opts.slice_direction,
         debug=opts.debug,
         mri_scheme=opts.mri_scheme,
         omp_nthreads=opts.omp_nthreads,

nirodents/data/artsBrainExtraction_precise_T2w.json

@@ -12,7 +12,7 @@
   "metric_weight": [
     1,
     1,
-    [0.5, 0.5]
+    [0.4, 0.6]
   ],
   "number_of_iterations": [
     [1000, 500, 250, 100],

nirodents/workflows/brainextraction.py

@@ -48,7 +48,6 @@ def init_rodent_brain_extraction_wf(
     arc=0.12,
     step=4,
     grid=(0, 4, 4),
-    slice_direction=1,
     debug=False,
     interim_checkpoints=True,
     mem_gb=3.0,
@@ -128,7 +127,7 @@ def init_rodent_brain_extraction_wf(
     norm_lap_tmpl = pe.Node(niu.Function(function=_trunc), name="norm_lap_tmpl")
     norm_lap_tmpl.inputs.out_max = 1.0
     norm_lap_tmpl.inputs.percentiles = (1, 99.99)
-    norm_lap_tmpl.inputs.clip_max = None
+    norm_lap_tmpl.inputs.clip_max = 99.9
     target_sigma = pe.Node(niu.Function(function=_lap_sigma),
                            name="target_sigma", run_without_submitting=True)
     lap_target = pe.Node(
@@ -137,7 +136,7 @@ def init_rodent_brain_extraction_wf(
     norm_lap_target = pe.Node(niu.Function(function=_trunc), name="norm_lap_target")
     norm_lap_target.inputs.out_max = 1.0
     norm_lap_target.inputs.percentiles = (1, 99.99)
-    norm_lap_target.inputs.clip_max = None
+    norm_lap_target.inputs.clip_max = 99.9
 
     # Set up initial spatial normalization
     ants_params = "testing" if debug else "precise"
@@ -167,8 +166,7 @@ def init_rodent_brain_extraction_wf(
     clip_tmpl.inputs.percentiles = (35.0, 90.0)
 
     # set INU bspline grid based on voxel size
-    init_bspline_grid = pe.Node(niu.Function(function=_bspline_distance), name="init_bspline_grid")
-    init_bspline_grid.inputs.slice_dir = slice_direction
+    bspline_grid = pe.Node(niu.Function(function=_bspline_grid), name="bspline_grid")
 
     # INU correction of the target image
     init_n4 = pe.Node(
@@ -200,8 +198,8 @@ def init_rodent_brain_extraction_wf(
     wf.connect([
         # Target image massaging
         (inputnode, clip_target, [(("in_files", _pop), "in_file")]),
-        (inputnode, init_bspline_grid, [(("in_files", _pop), "in_file")]),
-        (init_bspline_grid, init_n4, [("out", "args")]),
+        (inputnode, bspline_grid, [(("in_files", _pop), "in_file")]),
+        (bspline_grid, init_n4, [("out", "args")]),
         (clip_target, denoise, [("out", "input_image")]),
         (denoise, init_n4, [("output_image", "input_image")]),
         (init_n4, clip_inu, [("output_image", "in_file")]),
@@ -274,6 +272,7 @@ def init_rodent_brain_extraction_wf(
     # fmt: off
     wf.connect([
         (inputnode, map_brainmask, [(("in_files", _pop), "reference_image")]),
+        (bspline_grid, final_n4, [("out", "args")]),
         (denoise, final_n4, [("output_image", "input_image")]),
         # Project template's brainmask into subject space
         (norm, map_brainmask, [("reverse_transforms", "transforms"),
@@ -453,18 +452,18 @@ def _pop(in_files):
     return in_files
 
 
-def _bspline_distance(in_file, spacings=(8, 10, 8), slice_dir=1):
-    import numpy as np
+def _bspline_grid(in_file):
     import nibabel as nb
+    import numpy as np
 
     img = nb.load(in_file)
-    zooms = img.header.get_zooms()[:3]
-    zooms_round = [round(x, 3) for x in zooms]
-    extent = (np.array(img.shape[:3]) - 1) * zooms
-    if zooms_round.count(zooms_round[0]) != 3 and np.argmax(zooms) != slice_dir:
-        extent[np.argmax(zooms)], extent[slice_dir] = extent[slice_dir], extent[np.argmax(zooms)]
-    retval = [f"{v}" for v in np.ceil(extent / np.array(spacings)).astype(int)]
-    return f"-b [{'x'.join(retval)}]"
+    slices = img.header.get_data_shape()
+    # get slice ratio
+    ratio = [s / slices[np.argmax(slices)] for s in slices]
+    ratio_factor = ratio[np.argmax(ratio)] / ratio[np.argmin(ratio)]
+    # turn into integer resolution
+    mesh_res = [f"{round(i * ratio_factor)}" for i in ratio]
+    return f"-b [{'x'.join(mesh_res)}]"
 
 
 def _lap_sigma(in_file):
@@ -474,4 +473,4 @@ def _lap_sigma(in_file):
 
     img = nb.load(in_file)
     min_vox = np.amin(img.header.get_zooms())
-    return str(0.3508 * math.exp(1.4652 * min_vox))
+    return str(1.5 * min_vox ** 0.75)

setup.cfg

@@ -25,6 +25,8 @@ install_requires =
     nipype ~= 1.5.1
     niworkflows >= 1.3.2
     templateflow ~= 0.7.1
+    svgutils == 0.3.1
+
 test_requires =
     coverage < 5
     pytest >= 4.4