diff --git a/pyproject.toml b/pyproject.toml
index df3993a..79c2ea1 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -30,6 +30,7 @@ dependencies = [
     "scikit-image>=0.15.0",
     "scikit_learn>=1.3.0",
     "scipy>=1.8.0",
+    "typing_extensions >=4.12",
 ]
 dynamic = ["version"]
 
diff --git a/src/nifreeze/data/base.py b/src/nifreeze/data/base.py
index 4fee5a0..4cf47ff 100644
--- a/src/nifreeze/data/base.py
+++ b/src/nifreeze/data/base.py
@@ -27,7 +27,7 @@
 from collections import namedtuple
 from pathlib import Path
 from tempfile import mkdtemp
-from typing import Any, Generic, TypeVarTuple
+from typing import Any, Generic
 
 import attr
 import h5py
@@ -35,6 +35,7 @@
 import numpy as np
 from nibabel.spatialimages import SpatialHeader, SpatialImage
 from nitransforms.linear import Affine
+from typing_extensions import TypeVarTuple, Unpack
 
 from nifreeze.utils.ndimage import load_api
 
@@ -58,7 +59,7 @@ def _cmp(lh: Any, rh: Any) -> bool:
 
 
 @attr.s(slots=True)
-class BaseDataset(Generic[*Ts]):
+class BaseDataset(Generic[Unpack[Ts]]):
     """
     Base dataset representation structure.
 
@@ -99,13 +100,12 @@ def __len__(self) -> int:
 
         return self.dataobj.shape[-1]
 
-    def _getextra(self, idx: int | slice | tuple | np.ndarray) -> tuple[*Ts]:
-        # PY312: Default values for TypeVarTuples are not yet supported
+    def _getextra(self, idx: int | slice | tuple | np.ndarray) -> tuple[Unpack[Ts]]:
         return ()  # type: ignore[return-value]
 
     def __getitem__(
         self, idx: int | slice | tuple | np.ndarray
-    ) -> tuple[np.ndarray, np.ndarray | None, *Ts]:
+    ) -> tuple[np.ndarray, np.ndarray | None, Unpack[Ts]]:
         """
         Returns volume(s) and corresponding affine(s) through fancy indexing.
 
diff --git a/src/nifreeze/estimator.py b/src/nifreeze/estimator.py
index be1ed33..4c8a171 100644
--- a/src/nifreeze/estimator.py
+++ b/src/nifreeze/estimator.py
@@ -26,9 +26,10 @@
 
 from pathlib import Path
 from tempfile import TemporaryDirectory
-from typing import Self, TypeVar
+from typing import TypeVar
 
 from tqdm import tqdm
+from typing_extensions import Self
 
 from nifreeze.data.base import BaseDataset
 from nifreeze.model.base import BaseModel, ModelFactory
diff --git a/test/test_data_base.py b/test/test_data_base.py
index 4afa644..d8c65e2 100644
--- a/test/test_data_base.py
+++ b/test/test_data_base.py
@@ -53,25 +53,25 @@ def test_len(random_dataset: BaseDataset):
     assert len(random_dataset) == 5  # last dimension is 5 volumes
 
 
-def test_getitem_volume_index(random_dataset: BaseDataset[()]):
+def test_getitem_volume_index(random_dataset: BaseDataset):
     """
     Test that __getitem__ returns the correct (volume, affine) tuple.
 
     By default, motion_affines is None, so we expect to get None for the affine.
     """
-    # Single volume
-    volume0, aff0 = random_dataset[0]
+    # Single volume  # Note that the type ignore can be removed once we can use *Ts
+    volume0, aff0 = random_dataset[0]  # type: ignore[misc]  # PY310
     assert volume0.shape == (32, 32, 32)
     # No transforms have been applied yet, so there's no motion_affines array
     assert aff0 is None
 
     # Slice of volumes
-    volume_slice, aff_slice = random_dataset[2:4]
+    volume_slice, aff_slice = random_dataset[2:4]  # type: ignore[misc]  # PY310
     assert volume_slice.shape == (32, 32, 32, 2)
     assert aff_slice is None
 
 
-def test_set_transform(random_dataset: BaseDataset[()]):
+def test_set_transform(random_dataset: BaseDataset):
     """
     Test that calling set_transform changes the data and motion_affines.
     For simplicity, we'll apply an identity transform and check that motion_affines is updated.
@@ -83,7 +83,7 @@ def test_set_transform(random_dataset: BaseDataset[()]):
     random_dataset.set_transform(idx, affine, order=1)
 
     # Data shouldn't have changed (since transform is identity).
-    volume0, aff0 = random_dataset[idx]
+    volume0, aff0 = random_dataset[idx]  # type: ignore[misc]  # PY310
     assert np.allclose(data_before, volume0)
 
     # motion_affines should be created and match the transform matrix.