4103 enhances surface Dice to use subvoxel borders (#6681)

Fixes #4103

### Description
- considers spacing and subvoxel borders when computing surface Dice
- reimplemented
http://medicaldecathlon.com/files/Surface_distance_based_measures.ipynb
using pytorch API
- there's possibility of speed up once
#1332 is addressed


### Types of changes
<!--- Put an `x` in all the boxes that apply, and remove the not
applicable items -->
- [x] Non-breaking change (fix or new feature that would not break
existing functionality).
- [ ] Breaking change (fix or new feature that would cause existing
functionality to change).
- [x] New tests added to cover the changes.
- [ ] Integration tests passed locally by running `./runtests.sh -f -u
--net --coverage`.
- [x] Quick tests passed locally by running `./runtests.sh --quick
--unittests --disttests`.
- [x] In-line docstrings updated.
- [x] Documentation updated, tested `make html` command in the `docs/`
folder.

---------

Signed-off-by: Wenqi Li <[email protected]>

Author: wyli

monai/metrics/surface_dice.py

@@ -35,8 +35,9 @@ class SurfaceDiceMetric(CumulativeIterationMetric):
     Computes the Normalized Surface Dice (NSD) for each batch sample and class of
     predicted segmentations `y_pred` and corresponding reference segmentations `y` according to equation :eq:`nsd`.
     This implementation is based on https://arxiv.org/abs/2111.05408 and supports 2D and 3D images.
-    Be aware that the computation of boundaries is different from DeepMind's implementation
-    https://github.com/deepmind/surface-distance. In this implementation, the length/area of a segmentation boundary is
+    Be aware that by default (`use_subvoxels=False`), the computation of boundaries is different from DeepMind's
+    mplementation https://github.com/deepmind/surface-distance.
+    In this implementation, the length/area of a segmentation boundary is
     interpreted as the number of its edge pixels. In DeepMind's implementation, the length of a segmentation boundary
     depends on the local neighborhood (cf. https://arxiv.org/abs/1809.04430).
     This issue is discussed here: https://github.com/Project-MONAI/MONAI/issues/4103.
@@ -86,7 +87,7 @@ def _compute_tensor(self, y_pred: torch.Tensor, y: torch.Tensor, **kwargs: Any)
                 It must be a one-hot encoded, batch-first tensor [B,C,H,W] or [B,C,H,W,D].
             y: Reference segmentation.
                 It must be a one-hot encoded, batch-first tensor [B,C,H,W] or [B,C,H,W,D].
-            kwargs: additional parameters, e.g. ``spacing`` should be passed to correctly compute the metric.
+            kwargs: additional parameters: ``spacing`` should be passed to correctly compute the metric.
                 ``spacing``: spacing of pixel (or voxel). This parameter is relevant only
                 if ``distance_metric`` is set to ``"euclidean"``.
                 If a single number, isotropic spacing with that value is used for all images in the batch. If a sequence of numbers,
@@ -96,6 +97,8 @@ def _compute_tensor(self, y_pred: torch.Tensor, y: torch.Tensor, **kwargs: Any)
                 If inner sequence has length 1, isotropic spacing with that value is used for all images in the batch,
                 else the inner sequence length must be equal to the image dimensions. If ``None``, spacing of unity is used
                 for all images in batch. Defaults to ``None``.
+                use_subvoxels: Whether to use subvoxel distances. Defaults to ``False``.
+
 
         Returns:
             Pytorch Tensor of shape [B,C], containing the NSD values :math:`\operatorname {NSD}_{b,c}` for each batch
@@ -108,6 +111,7 @@ def _compute_tensor(self, y_pred: torch.Tensor, y: torch.Tensor, **kwargs: Any)
             include_background=self.include_background,
             distance_metric=self.distance_metric,
             spacing=kwargs.get("spacing"),
+            use_subvoxels=kwargs.get("use_subvoxels", False),
         )
 
     def aggregate(
@@ -141,13 +145,14 @@ def compute_surface_dice(
     include_background: bool = False,
     distance_metric: str = "euclidean",
     spacing: int | float | np.ndarray | Sequence[int | float | np.ndarray | Sequence[int | float]] | None = None,
+    use_subvoxels: bool = False,
 ) -> torch.Tensor:
     r"""
     This function computes the (Normalized) Surface Dice (NSD) between the two tensors `y_pred` (referred to as
     :math:`\hat{Y}`) and `y` (referred to as :math:`Y`). This metric determines which fraction of a segmentation
     boundary is correctly predicted. A boundary element is considered correctly predicted if the closest distance to the
-    reference boundary is smaller than or equal to the specified threshold related to the acceptable amount of deviation in
-    pixels. The NSD is bounded between 0 and 1.
+    reference boundary is smaller than or equal to the specified threshold related to the acceptable amount of deviation
+    in pixels. The NSD is bounded between 0 and 1.
 
     This implementation supports multi-class tasks with an individual threshold :math:`\tau_c` for each class :math:`c`.
     The class-specific NSD for batch index :math:`b`, :math:`\operatorname {NSD}_{b,c}`, is computed using the function:
@@ -159,24 +164,23 @@ def compute_surface_dice(
         :label: nsd
 
     with :math:`\mathcal{D}_{Y_{b,c}}` and :math:`\mathcal{D}_{\hat{Y}_{b,c}}` being two sets of nearest-neighbor
-    distances. :math:`\mathcal{D}_{Y_{b,c}}` is computed from the predicted segmentation boundary towards the reference segmentation
-    boundary and vice-versa for :math:`\mathcal{D}_{\hat{Y}_{b,c}}`. :math:`\mathcal{D}_{Y_{b,c}}^{'}` and
+    distances. :math:`\mathcal{D}_{Y_{b,c}}` is computed from the predicted segmentation boundary towards the reference
+    segmentation boundary and vice-versa for :math:`\mathcal{D}_{\hat{Y}_{b,c}}`. :math:`\mathcal{D}_{Y_{b,c}}^{'}` and
     :math:`\mathcal{D}_{\hat{Y}_{b,c}}^{'}` refer to the subsets of distances that are smaller or equal to the
     acceptable distance :math:`\tau_c`:
 
     .. math::
         \mathcal{D}_{Y_{b,c}}^{'} = \{ d \in \mathcal{D}_{Y_{b,c}} \, | \, d \leq \tau_c \}.
 
 
-    In the case of a class neither being present in the predicted segmentation, nor in the reference segmentation, a nan value
-    will be returned for this class. In the case of a class being present in only one of predicted segmentation or
-    reference segmentation, the class NSD will be 0.
+    In the case of a class neither being present in the predicted segmentation, nor in the reference segmentation,
+    a nan value will be returned for this class. In the case of a class being present in only one of predicted
+    segmentation or reference segmentation, the class NSD will be 0.
 
     This implementation is based on https://arxiv.org/abs/2111.05408 and supports 2D and 3D images.
-    Be aware that the computation of boundaries is different from DeepMind's implementation
-    https://github.com/deepmind/surface-distance. In this implementation, the length of a segmentation boundary is
-    interpreted as the number of its edge pixels. In DeepMind's implementation, the length of a segmentation boundary
-    depends on the local neighborhood (cf. https://arxiv.org/abs/1809.04430).
+    The computation of boundaries follows DeepMind's implementation
+    https://github.com/deepmind/surface-distance when `use_subvoxels=True`; Otherwise the length of a segmentation
+    boundary is interpreted as the number of its edge pixels.
 
     Args:
         y_pred: Predicted segmentation, typically segmentation model output.
@@ -198,6 +202,7 @@ def compute_surface_dice(
             If inner sequence has length 1, isotropic spacing with that value is used for all images in the batch,
             else the inner sequence length must be equal to the image dimensions. If ``None``, spacing of unity is used
             for all images in batch. Defaults to ``None``.
+        use_subvoxels: Whether to use subvoxel distances. Defaults to ``False``.
 
     Raises:
         ValueError: If `y_pred` and/or `y` are not PyTorch tensors.
@@ -227,11 +232,6 @@ def compute_surface_dice(
             f"y_pred and y should have same shape, but instead, shapes are {y_pred.shape} (y_pred) and {y.shape} (y)."
         )
 
-    if not torch.all(y_pred.byte() == y_pred) or not torch.all(y.byte() == y):
-        raise ValueError("y_pred and y should be binarized tensors (e.g. torch.int64).")
-    if torch.any(y_pred > 1) or torch.any(y > 1):
-        raise ValueError("y_pred and y should be one-hot encoded.")
-
     y = y.float()
     y_pred = y_pred.float()
 
@@ -254,24 +254,37 @@ def compute_surface_dice(
     spacing_list = prepare_spacing(spacing=spacing, batch_size=batch_size, img_dim=img_dim)
 
     for b, c in np.ndindex(batch_size, n_class):
-        (edges_pred, edges_gt) = get_mask_edges(y_pred[b, c], y[b, c], crop=False)
+        if not use_subvoxels:
+            (edges_pred, edges_gt) = get_mask_edges(y_pred[b, c], y[b, c], crop=True)
+            distances_pred_gt = get_surface_distance(
+                edges_pred, edges_gt, distance_metric=distance_metric, spacing=spacing_list[b]
+            )
+            distances_gt_pred = get_surface_distance(
+                edges_gt, edges_pred, distance_metric=distance_metric, spacing=spacing_list[b]
+            )
+
+            boundary_complete = len(distances_pred_gt) + len(distances_gt_pred)
+            boundary_correct = np.sum(distances_pred_gt <= class_thresholds[c]) + np.sum(
+                distances_gt_pred <= class_thresholds[c]
+            )
+        else:
+            _spacing = spacing_list[b] if spacing_list[b] is not None else [1] * img_dim
+            areas_pred: np.ndarray
+            areas_gt: np.ndarray
+            edges_pred, edges_gt, areas_pred, areas_gt = get_mask_edges(  # type: ignore
+                y_pred[b, c], y[b, c], crop=True, spacing=_spacing  # type: ignore
+            )
+            dist_pred_to_gt = get_surface_distance(edges_pred, edges_gt, distance_metric, spacing=spacing_list[b])
+            dist_gt_to_pred = get_surface_distance(edges_gt, edges_pred, distance_metric, spacing=spacing_list[b])
+            areas_gt, areas_pred = areas_gt[edges_gt], areas_pred[edges_pred]
+            boundary_complete = areas_gt.sum() + areas_pred.sum()
+            gt_true = areas_gt[dist_gt_to_pred <= class_thresholds[c]].sum() if len(areas_gt) > 0 else 0.0
+            pred_true = areas_pred[dist_pred_to_gt <= class_thresholds[c]].sum() if len(areas_pred) > 0 else 0.0
+            boundary_correct = gt_true + pred_true
         if not np.any(edges_gt):
             warnings.warn(f"the ground truth of class {c} is all 0, this may result in nan/inf distance.")
         if not np.any(edges_pred):
             warnings.warn(f"the prediction of class {c} is all 0, this may result in nan/inf distance.")
-
-        distances_pred_gt = get_surface_distance(
-            edges_pred, edges_gt, distance_metric=distance_metric, spacing=spacing_list[b]
-        )
-        distances_gt_pred = get_surface_distance(
-            edges_gt, edges_pred, distance_metric=distance_metric, spacing=spacing_list[b]
-        )
-
-        boundary_complete = len(distances_pred_gt) + len(distances_gt_pred)
-        boundary_correct = np.sum(distances_pred_gt <= class_thresholds[c]) + np.sum(
-            distances_gt_pred <= class_thresholds[c]
-        )
-
         if boundary_complete == 0:
             # the class is neither present in the prediction, nor in the reference segmentation
             nsd[b, c] = np.nan