ENH: Execute FSL FAST only with --fs-no-reconall

This PR restructures the anatomical workflow so that the ``recon-all``
based graph does not execute FSL FAST, and rather projects the
``aseg`` segmentation to a 3-tissue brain segmentation with labels
according to BIDS.

This should provide the confounds workflow of *fMRIPrep* with smoother
(and probably more accurate) regions (e.g., for aCompCor calculations).
On the long run, we hope this will add some extra run-to-run
reproducibility to the aCompCor calculations.

Resolves: #170

smriprep/workflows/anatomical.py

@@ -320,22 +320,6 @@ def _check_img(img):
             omp_nthreads=omp_nthreads,
             normalization_quality='precise' if not debug else 'testing')
 
-    # 3. Brain tissue segmentation - FAST produces: 0 (bg), 1 (wm), 2 (csf), 3 (gm)
-    t1w_dseg = pe.Node(fsl.FAST(segments=True, no_bias=True, probability_maps=True),
-                       name='t1w_dseg', mem_gb=3)
-
-    # Change LookUp Table - BIDS wants: 0 (bg), 1 (gm), 2 (wm), 3 (csf)
-    lut_t1w_dseg = pe.Node(niu.Function(function=_apply_bids_lut),
-                           name='lut_t1w_dseg')
-    lut_t1w_dseg.inputs.lut = (0, 3, 1, 2)  # Maps: 0 -> 0, 3 -> 1, 1 -> 2, 2 -> 3.
-
-    workflow.connect([
-        (buffernode, t1w_dseg, [('t1w_brain', 'in_files')]),
-        (t1w_dseg, lut_t1w_dseg, [('tissue_class_map', 'in_dseg')]),
-        (t1w_dseg, outputnode, [('probability_maps', 't1w_tpms')]),
-        (lut_t1w_dseg, outputnode, [('out', 't1w_dseg')]),
-    ])
-
     # 4. Spatial normalization
     anat_norm_wf = init_anat_norm_wf(
         debug=debug,
@@ -363,10 +347,6 @@ def _check_img(img):
         (brain_extraction_wf, anat_norm_wf, [
             (('outputnode.bias_corrected', _pop), 'inputnode.moving_image')]),
         (buffernode, anat_norm_wf, [('t1w_mask', 'inputnode.moving_mask')]),
-        (lut_t1w_dseg, anat_norm_wf, [
-            ('out', 'inputnode.moving_segmentation')]),
-        (t1w_dseg, anat_norm_wf, [
-            ('probability_maps', 'inputnode.moving_tpms')]),
         (anat_norm_wf, outputnode, [
             ('poutputnode.standardized', 'std_preproc'),
             ('poutputnode.std_mask', 'std_mask'),
@@ -378,6 +358,16 @@ def _check_img(img):
         ]),
     ])
 
+    # Change LookUp Table - BIDS wants: 0 (bg), 1 (gm), 2 (wm), 3 (csf)
+    lut_t1w_dseg = pe.Node(niu.Function(function=_apply_bids_lut),
+                           name='lut_t1w_dseg')
+
+    workflow.connect([
+        (lut_t1w_dseg, anat_norm_wf, [
+            ('out', 'inputnode.moving_segmentation')]),
+        (lut_t1w_dseg, outputnode, [('out', 't1w_dseg')]),
+    ])
+
     # Connect reportlets
     workflow.connect([
         (inputnode, anat_reports_wf, [
@@ -423,13 +413,28 @@ def _check_img(img):
     ])
 
     if not freesurfer:  # Flag --fs-no-reconall is set - return
+        # Brain tissue segmentation - FAST produces: 0 (bg), 1 (wm), 2 (csf), 3 (gm)
+        t1w_dseg = pe.Node(fsl.FAST(segments=True, no_bias=True, probability_maps=True),
+                           name='t1w_dseg', mem_gb=3)
+        lut_t1w_dseg.inputs.lut = (0, 3, 1, 2)  # Maps: 0 -> 0, 3 -> 1, 1 -> 2, 2 -> 3.
+
         workflow.connect([
             (brain_extraction_wf, buffernode, [
                 (('outputnode.out_file', _pop), 't1w_brain'),
                 ('outputnode.out_mask', 't1w_mask')]),
+            (buffernode, t1w_dseg, [('t1w_brain', 'in_files')]),
+            (t1w_dseg, lut_t1w_dseg, [('tissue_class_map', 'in_dseg')]),
+            (t1w_dseg, anat_norm_wf, [
+                ('probability_maps', 'inputnode.moving_tpms')]),
+            (t1w_dseg, outputnode, [('probability_maps', 't1w_tpms')]),
         ])
         return workflow
 
+    # Map FS' aseg labels onto three-tissue segmentation
+    lut_t1w_dseg.inputs.lut = _aseg_two_three()
+    split_seg = pe.Node(niu.Function(function=_split_segments),
+                        name='split_seg')
+
     # 5. Surface reconstruction (--fs-no-reconall not set)
     surface_recon_wf = init_surface_recon_wf(name='surface_recon_wf',
                                              omp_nthreads=omp_nthreads, hires=hires)
@@ -449,6 +454,12 @@ def _check_img(img):
             (('outputnode.bias_corrected', _pop), 'in_file')]),
         (surface_recon_wf, applyrefined, [
             ('outputnode.out_brainmask', 'mask_file')]),
+        (surface_recon_wf, lut_t1w_dseg, [
+            ('outputnode.out_aseg', 'in_dseg')]),
+        (lut_t1w_dseg, split_seg, [('out', 'in_file')]),
+        (split_seg, anat_norm_wf, [
+            ('out', 'inputnode.moving_tpms')]),
+        (split_seg, outputnode, [('out', 't1w_tpms')]),
         (surface_recon_wf, outputnode, [
             ('outputnode.subjects_dir', 'subjects_dir'),
             ('outputnode.subject_id', 'subject_id'),
@@ -620,3 +631,55 @@ def _pop(inlist):
     if isinstance(inlist, (list, tuple)):
         return inlist[0]
     return inlist
+
+
+def _aseg_two_three():
+    import numpy as np
+    # Base struct
+    aseg_lut = np.zeros((256,), dtype="int")
+    # GM
+    aseg_lut[3] = 1
+    aseg_lut[8:14] = 1
+    aseg_lut[17:21] = 1
+    aseg_lut[26:40] = 1
+    aseg_lut[42] = 1
+    aseg_lut[47:73] = 1
+
+    # CSF
+    aseg_lut[4:6] = 3
+    aseg_lut[14:16] = 3
+    aseg_lut[24] = 3
+    aseg_lut[43:45] = 3
+    aseg_lut[72] = 3
+
+    # WM
+    aseg_lut[2] = 2
+    aseg_lut[7] = 2
+    aseg_lut[16] = 2
+    aseg_lut[28] = 2
+    aseg_lut[41] = 2
+    aseg_lut[46] = 2
+    aseg_lut[60] = 2
+    aseg_lut[77:80] = 2
+    aseg_lut[250:256] = 2
+    return tuple(aseg_lut)
+
+
+def _split_segments(in_file):
+    from pathlib import Path
+    import numpy as np
+    import nibabel as nb
+
+    segimg = nb.load(in_file)
+    data = np.int16(segimg.dataobj)
+    hdr = segimg.header.copy()
+    hdr.set_data_dtype('uint8')
+
+    out_files = []
+    for i, label in enumerate(("CSF", "WM", "GM")):
+        out_fname = str((Path() / f"aseg_label-{label}_mask.nii.gz").absolute())
+        segment = (data == i + 1).astype('uint8')
+        segimg.__class__(segment, segimg.affine, hdr).to_filename(out_fname)
+        out_files.append(out_fname)
+
+    return out_files