24.1.x

deepprep/nextflow/bin/bold_apply_transform_chain.py

@@ -7,12 +7,19 @@
 import numpy as np
 from pathlib import Path
 from scipy import ndimage as ndi
+from scipy.sparse import hstack as sparse_hstack
 import argparse
 
 from nitransforms.io import get_linear_factory
 import templateflow.api as tflow
 from multiprocessing import Pool
 from bids import BIDSLayout
+import json
+
+from sdcflows.utils.tools import ensure_positive_cosines
+from sdcflows.transform import grid_bspline_weights
+import nitransforms as nt
+
 
 
 def get_preproc_file(subject_id, bids_preproc, bold_orig_file, update_entities):
@@ -32,12 +39,12 @@ def get_preproc_file(subject_id, bids_preproc, bold_orig_file, update_entities):
     return Path(bold_t1w_file)
 
 
-def affine_to_3x3(itk):
-    matrix = itk[:3, :3]
-    translation = itk[:3, -1:]
+def affine_to_3x3(mtx):
+    matrix = mtx[:3, :3]
+    translation = mtx[:3, -1:]
     return matrix, translation
 
-def apply_hmc_pool(frame, warped_mesh_frame, matrix_frame, ras2vox_A, ras2vox_b, bold_orig, fixed, bold_orig_header, transform_save_path):
+def apply_hmc_pool(frame, warped_mesh_frame, matrix_frame, ras2vox_A, ras2vox_b, bold_orig, fixed, bold_orig_header, transform_save_path, bold_sdc, pe_info, vsm):
     # bold_orig = nib.load(bold_file)
     bold_orig_values = bold_orig.slicer[..., frame:frame+1].get_fdata()[..., 0]
 
@@ -47,6 +54,9 @@ def apply_hmc_pool(frame, warped_mesh_frame, matrix_frame, ras2vox_A, ras2vox_b,
     warped_mesh_frame = ras2vox_A @ warped_mesh_frame + ras2vox_b
     warped_mesh_frame = warped_mesh_frame.reshape(3, *fixed.shape)
 
+    if bold_sdc:
+        warped_mesh_frame[pe_info[0][0], ...] += vsm
+
     # interp values
     output = np.zeros(
         list(fixed.shape),
@@ -62,6 +72,11 @@ def apply_hmc_pool(frame, warped_mesh_frame, matrix_frame, ras2vox_A, ras2vox_b,
         cval=0.0,
         prefilter=True,
     )
+
+    if bold_sdc:
+        # jacobian
+        result *= 1 + np.gradient(vsm, axis=pe_info[0][0])
+
     result = result[..., np.newaxis]
     nib.save(nib.Nifti1Image(result, affine=fixed.affine, header=bold_orig_header),
          f'{transform_save_path}/t{str(frame).zfill(5)}.nii.gz')
@@ -73,12 +88,45 @@ def concat_frames(transform_save_path, output_path, boldref_path, t1_json):
     cmd = f'mri_concat --i {transform_save_path}/* --o {output_path}'
     os.system(cmd)
 
-    # copy the first frame as boldref
-    shutil.copy(in_files[0], boldref_path)
-
     # generate .json, it is consistent with T1w.json
     boldref_json_path = str(output_path).replace('.nii.gz', '.json')
-    shutil.copy(t1_json, boldref_json_path)
+    try:
+        # copy the first frame as boldref
+        shutil.copy(in_files[0], boldref_path)
+        shutil.copy(t1_json, boldref_json_path)
+    except:
+        pass
+
+
+def apply_fieldmap2std(in_coeffs, target_ref_file, fmap_ref_file, transforms):
+    coefficients = [nib.load(in_coeffs)]
+    target = nib.load(target_ref_file)
+    fmapref = nib.load(fmap_ref_file)
+
+    warp_matrix = nib.load(transforms[0]).get_fdata()
+    warp_affine = nib.load(transforms[0]).affine
+    transform_chain = nt.TransformChain(nt.DenseFieldTransform(nib.Nifti1Image(warp_matrix, warp_affine)))
+    xfm = nt.linear.load(transforms[1])
+    transform_chain += xfm
+    xfm = nt.linear.load(transforms[2])
+    transform_chain += xfm
+    reference, _ = ensure_positive_cosines(fmapref)
+    colmat = sparse_hstack(
+        [grid_bspline_weights(reference, level) for level in coefficients]
+    ).tocsr()
+    coefficients = np.hstack(
+        [level.get_fdata(dtype='float32').reshape(-1) for level in coefficients]
+    )
+    fmap_img = nib.Nifti1Image(
+        np.reshape(colmat @ coefficients, reference.shape[:3]),
+        reference.affine,
+    )
+    fmap_img = transform_chain.apply(fmap_img, reference=target)
+    fmap_img.header.set_intent('estimate', name='fieldmap Hz')
+    fmap_img.header.set_data_dtype('float32')
+    fmap_img.header['cal_max'] = max((abs(fmap_img.dataobj.min()), fmap_img.dataobj.max()))
+    fmap_img.header['cal_min'] = -fmap_img.header['cal_max']
+    return fmap_img
 
 
 if __name__ == '__main__':
@@ -106,6 +154,8 @@ def concat_frames(transform_save_path, output_path, boldref_path, t1_json):
     parser.add_argument("--template_space", required=True)
     parser.add_argument("--template_resolution", required=True)
     parser.add_argument("--nonlinear_file", required=True)
+    parser.add_argument("--bold_sdc", required=True)
+    parser.add_argument("--task_id", required=True)
     parser.add_argument("--reference", required=False, default=None)
     parser.add_argument("--moving", required=False, default=None)
     args = parser.parse_args()
@@ -137,6 +187,33 @@ def concat_frames(transform_save_path, output_path, boldref_path, t1_json):
     output_path = Path(coreg_xfm.parent) / f'{args.bold_id}_space-{args.template_space}_res-{args.template_resolution}_desc-preproc_bold.nii.gz'
     boldref_path = Path(coreg_xfm.parent) / f'{args.bold_id}_space-{args.template_space}_res-{args.template_resolution}_boldref.nii.gz'
 
+    # load fieldmap info
+    if args.bold_sdc:
+        update_entities = {'suffix': 'bold', 'extension': '.json'}
+        bold_json = get_preproc_file(args.subject_id, args.bids_dir, bold_file, update_entities)
+        with open(str(bold_json)) as f:
+            bold_info = json.load(f)
+        pe_dir = bold_info["PhaseEncodingDirection"]
+        ro_time = bold_info["TotalReadoutTime"]
+
+        fmap_base_dir = Path(args.work_dir) / 'bold_preprocess' /f'{args.subject_id}_wf' / f'{args.task_id}_wf'
+        fieldmap_id_txt_path = fmap_base_dir / 'fieldmap_id.txt'
+        with open(str(fieldmap_id_txt_path)) as f:
+            fieldmap_id_info = json.load(f)
+        fieldmap_id = fieldmap_id_info.get(bold_file)
+
+        coeff_dir = fmap_base_dir.parent / 'fmap_preproc_wf' / f'wf_{fieldmap_id}' / 'fix_coeff'
+        in_coeff = sorted(coeff_dir.glob("*_fieldcoef_fixed.nii.gz"))[0]
+
+
+        fmap_ref_file = str(fmap_base_dir.parent / 'fmap_preproc_wf' / f'wf_{fieldmap_id}' / 'brainextraction_wf' / 'clipper_post' / 'clipped.nii.gz')
+
+        # get the coreg.xfm
+        update_entities = {'mode': 'image', 'suffix': 'xfm', 'extension': '.txt'}
+        fieldmap_xfm = get_preproc_file(args.subject_id, args.bold_preprocess_dir, bold_file, update_entities)
+        transforms = [nonlinear_file, coreg_xfm, fieldmap_xfm]
+        fieldmap = apply_fieldmap2std(in_coeff, fixed_file, fmap_ref_file, transforms)
+
     # Load the fixed file
     fixed = nib.load(fixed_file)
     vox2ras = fixed.affine
@@ -181,9 +258,30 @@ def concat_frames(transform_save_path, output_path, boldref_path, t1_json):
     ras2vox_bold = np.linalg.inv(bold_orig.affine)
     ras2vox_A, ras2vox_b = affine_to_3x3(ras2vox_bold)
 
+    # apply fieldmap if available
+    if args.bold_sdc:
+        nvols = bold_orig.shape[3] if bold_orig.ndim > 3 else 1
+
+        if pe_dir and ro_time:
+            pe_axis = "ijk".index(pe_dir[0])
+            pe_flip = pe_dir.endswith("-")
+
+            # Nitransforms displacements are positive
+            source, axcodes = ensure_positive_cosines(bold_orig)
+            axis_flip = axcodes[pe_axis] in "LPI"
+
+            pe_info = [(pe_axis, -ro_time if (axis_flip ^ pe_flip) else ro_time)] * nvols
+
+        fmap_hz = fieldmap.get_fdata(dtype='f4')
+        vsm = fmap_hz * pe_info[0][1]
+
+    else:
+        pe_info = None
+        vsm = None
+
     args_apply_hmc = []
     for i in range(matrix.shape[0]):
-        args_apply_hmc.append([int(i), warped_mesh, matrix[i], ras2vox_A, ras2vox_b, bold_orig, fixed, bold_orig_header, transform_save_path])
+        args_apply_hmc.append([int(i), warped_mesh, matrix[i], ras2vox_A, ras2vox_b, bold_orig, fixed, bold_orig_header, transform_save_path, args.bold_sdc, pe_info, vsm])
     pool = Pool(10)
     pool.starmap(apply_hmc_pool, args_apply_hmc)
     pool.close()

deepprep/nextflow/bin/bold_preprocess.py

@@ -3,6 +3,7 @@
 import os
 import shutil
 from pathlib import Path
+import json
 
 from fmriprep.workflows.fieldmap import init_single_subject_fieldmap_wf
 from fmriprep.workflows.bold.base import init_bold_wf
@@ -235,6 +236,10 @@ def get_bold_func_path(subject_id, bids_preproc, bold_orig_file):
         fmap_base_dir = Path(config.execution.work_dir) / f'{subject_id}_wf'
         base_dir = fmap_base_dir / f'{args.task_id[0]}_wf'
         base_dir.mkdir(parents=True, exist_ok=True)
+        # save fieldmap id
+        fieldmap_id_txt = base_dir / 'fieldmap_id.txt'
+        with open(fieldmap_id_txt, 'w') as f:
+            json.dump(estimator_map, f)
         workflow.connect([
             (inputnode, bold_wf, [
                 ("t1w_preproc", "inputnode.t1w_preproc"),

deepprep/nextflow/deepprep.nf

@@ -2102,12 +2102,14 @@ process bold_transform_chain {
     tuple(val(subject_id), val(bold_id), val(trans), path(subject_boldfile_txt_bold))
     val(template_space)
     val(template_resolution)
+    val(bold_sdc)
 
     output:
     tuple(val(subject_id), val(bold_id), val("${bold_preprocess_path}/${subject_id}/func/${bold_id}_space-${template_space}_res-${template_resolution}_desc-preproc_bold.nii.gz"))
     tuple(val(subject_id), val(bold_id), val("${bold_preprocess_path}/${subject_id}/func/${bold_id}_space-${template_space}_res-${template_resolution}_boldref.nii.gz"))
 
     script:
+    task_id = bold_id.split('task-')[1].split('_')[0]
     script_py = "bold_apply_transform_chain.py"
 
     """
@@ -2120,7 +2122,9 @@ process bold_transform_chain {
     --subject_boldfile_txt_bold ${subject_boldfile_txt_bold} \
     --template_space ${template_space} \
     --template_resolution ${template_resolution} \
-    --nonlinear_file ${trans}
+    --nonlinear_file ${trans} \
+    --bold_sdc ${bold_sdc} \
+    --task_id ${task_id}
     """
 }
 
@@ -2976,7 +2980,7 @@ workflow bold_wf {
         (t1_norigid_nii, norm_norigid_nii, trans) = bold_synthmorph_joint(subjects_dir, bold_preprocess_path, synthmorph_home, bold_synthmorph_joint_input, synthmorph_model_path, template_space, device, gpu_lock)
 
         bold_transform_chain_input = subject_id_boldfile_id.groupTuple(sort: true).join(trans, by:[0]).transpose().join(subject_boldfile_txt_bold_pre_process, by: [0, 1])
-        (preproc_bold, boldref_file) = bold_transform_chain(bids_dir, bold_preprocess_path, work_dir, bold_transform_chain_input, template_space, template_resolution)
+        (preproc_bold, boldref_file) = bold_transform_chain(bids_dir, bold_preprocess_path, work_dir, bold_transform_chain_input, template_space, template_resolution, bold_sdc)
     }
 
     do_bold_qc = 'TRUE'