nipy · oesteban · Jul 20, 2022 · Jul 17, 2025 · Jul 17, 2025 · Jul 17, 2025
diff --git a/nitransforms/resampling.py b/nitransforms/resampling.py
@@ -10,6 +10,7 @@
 
 import asyncio
 from os import cpu_count
+from contextlib import suppress
 from functools import partial
 from pathlib import Path
 from typing import Callable, TypeVar, Union
@@ -108,12 +109,17 @@ async def _apply_serial(
     semaphore = asyncio.Semaphore(max_concurrent)
 
     for t in range(n_resamplings):
-        xfm_t = transform if (n_resamplings == 1 or transform.ndim < 4) else transform[t]
+        xfm_t = (
+            transform if (n_resamplings == 1 or transform.ndim < 4) else transform[t]
+        )
 
-        if targets is None:
-            targets = ImageGrid(spatialimage).index(  # data should be an image
+        targets_t = (
+            ImageGrid(spatialimage).index(
                 _as_homogeneous(xfm_t.map(ref_ndcoords), dim=ref_ndim)
             )
+            if targets is None
+            else targets[t, ...]
+        )
 
         data_t = (
             data
@@ -127,7 +133,7 @@ async def _apply_serial(
                     partial(
                         ndi.map_coordinates,
                         data_t,
-                        targets,
+                        targets_t,
                         output=output[..., t],
                         order=order,
                         mode=mode,
@@ -255,11 +261,22 @@ def apply(
                 dim=_ref.ndim,
             )
         )
-    elif xfm_nvols == 1:
-        targets = ImageGrid(spatialimage).index(  # data should be an image
-            _as_homogeneous(transform.map(ref_ndcoords), dim=_ref.ndim)
+    else:
+        # Targets' shape is (Nt, 3, Nv) with Nv = Num. voxels, Nt = Num. timepoints.
+        targets = (
+            ImageGrid(spatialimage).index(
+                _as_homogeneous(transform.map(ref_ndcoords), dim=_ref.ndim)
+            )
+            if targets is None
+            else targets
         )
 
+    if targets.ndim == 3:
+        targets = np.rollaxis(targets, targets.ndim - 1, 0)
+    else:
+        assert targets.ndim == 2
+        targets = targets[np.newaxis, ...]
+
     if serialize_4d:
         data = (
             np.asanyarray(spatialimage.dataobj, dtype=input_dtype)
@@ -294,17 +311,24 @@ def apply(
     else:
         data = np.asanyarray(spatialimage.dataobj, dtype=input_dtype)
 
-        if targets is None:
-            targets = ImageGrid(spatialimage).index(  # data should be an image
-                _as_homogeneous(transform.map(ref_ndcoords), dim=_ref.ndim)
-            )
-
+        if data_nvols == 1 and xfm_nvols == 1:
+            targets = np.squeeze(targets)
+            assert targets.ndim == 2
         # Cast 3D data into 4D if 4D nonsequential transform
-        if data_nvols == 1 and xfm_nvols > 1:
+        elif data_nvols == 1 and xfm_nvols > 1:
             data = data[..., np.newaxis]
 
-        if transform.ndim == 4:
-            targets = _as_homogeneous(targets.reshape(-2, targets.shape[0])).T
+        if xfm_nvols > 1:
+            assert targets.ndim == 3
+            n_time, n_dim, n_vox = targets.shape
+            # Reshape to (3, n_time x n_vox)
+            ijk_targets = np.rollaxis(targets, 0, 2).reshape((n_dim, -1))
+            time_row = np.repeat(np.arange(n_time), n_vox)[None, :]
+
+            # Now targets is (4, n_vox x n_time), with indexes (t, i, j, k)
+            # t is the slowest-changing axis, so we put it first
+            targets = np.vstack((time_row, ijk_targets))
+            data = np.rollaxis(data, data.ndim - 1, 0)
 
         resampled = ndi.map_coordinates(
             data,
@@ -323,11 +347,19 @@ def apply(
         )
         hdr.set_data_dtype(output_dtype or spatialimage.header.get_data_dtype())
 
-        moved = spatialimage.__class__(
-            resampled.reshape(_ref.shape if n_resamplings == 1 else _ref.shape + (-1,)),
-            _ref.affine,
-            hdr,
-        )
+        if serialize_4d:
+            resampled = resampled.reshape(
+                _ref.shape
+                if n_resamplings == 1
+                else _ref.shape + (resampled.shape[-1],)
+            )
+        else:
+            resampled = resampled.reshape((-1, *_ref.shape))
+            resampled = np.rollaxis(resampled, 0, resampled.ndim)
+            with suppress(ValueError):
+                resampled = np.squeeze(resampled, axis=3)
+
+        moved = spatialimage.__class__(resampled, _ref.affine, hdr)
         return moved
 
     output_dtype = output_dtype or input_dtype

diff --git a/nitransforms/tests/test_resampling.py b/nitransforms/tests/test_resampling.py
@@ -363,3 +363,28 @@ def test_LinearTransformsMapping_apply(
             reference=testdata_path / "sbref.nii.gz",
             serialize_nvols=2 if serialize_4d else np.inf,
         )
+
+
+@pytest.mark.parametrize("serialize_4d", [True, False])
+def test_apply_4d(serialize_4d):
+    """Regression test for per-volume transforms with serialized resampling."""
+    nvols = 9
+    shape = (10, 5, 5)
+    base = np.zeros(shape, dtype=np.float32)
+    base[9, 2, 2] = 1
+    img = nb.Nifti1Image(np.stack([base] * nvols, axis=-1), np.eye(4))
+
+    transforms = []
+    for i in range(nvols):
+        mat = np.eye(4)
+        mat[0, 3] = i
+        transforms.append(nitl.Affine(mat))
+
+    extraparams = {} if serialize_4d else {"serialize_nvols": nvols + 1}
+
+    xfm = nitl.LinearTransformsMapping(transforms, reference=img)
+
+    moved = apply(xfm, img, order=0, **extraparams)
+    data = np.asanyarray(moved.dataobj)
+    idxs = [tuple(np.argwhere(data[..., i])[0]) for i in range(nvols)]
+    assert idxs == [(9 - i, 2, 2) for i in range(nvols)]