fix: address review feedback — fix class_index, eliminate double-masking, add docs

Signed-off-by: Rusheel Sharma <rusheelhere@gmail.com>

Author: Rusheel86

monai/losses/unified_focal_loss.py

@@ -225,12 +225,8 @@ def __init__(
         self.gamma = gamma
         self.delta = delta
         self.weight: float = weight
-        self.asy_focal_loss = AsymmetricFocalLoss(
-            to_onehot_y=False, gamma=self.gamma, delta=self.delta, ignore_index=ignore_index
-        )
-        self.asy_focal_tversky_loss = AsymmetricFocalTverskyLoss(
-            to_onehot_y=False, gamma=self.gamma, delta=self.delta, ignore_index=ignore_index
-        )
+        self.asy_focal_loss = AsymmetricFocalLoss(to_onehot_y=False, gamma=self.gamma, delta=self.delta)
+        self.asy_focal_tversky_loss = AsymmetricFocalTverskyLoss(to_onehot_y=False, gamma=self.gamma, delta=self.delta)
         self.ignore_index = ignore_index
 
     def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
@@ -283,14 +279,11 @@ def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
 
         mask = create_ignore_mask(original_y_true, self.ignore_index)
 
-        use_mask = self.ignore_index is not None and (self.ignore_index < 0 or self.ignore_index >= self.num_classes)
-        if use_mask:
-            y_pred_masked, y_true_masked = mask_loss_inputs(y_pred, y_true, self.ignore_index, mask=mask)
-        else:
-            y_pred_masked, y_true_masked = y_pred, y_true
+        if self.ignore_index is not None:
+            y_pred, y_true = mask_loss_inputs(y_pred, y_true, self.ignore_index, mask=mask)
 
-        asy_focal_loss = self.asy_focal_loss(y_pred_masked, y_true_masked)
-        asy_focal_tversky_loss = self.asy_focal_tversky_loss(y_pred_masked, y_true_masked)
+        asy_focal_loss = self.asy_focal_loss(y_pred, y_true)
+        asy_focal_tversky_loss = self.asy_focal_tversky_loss(y_pred, y_true)
 
         loss: torch.Tensor = self.weight * asy_focal_loss + (1 - self.weight) * asy_focal_tversky_loss
 

monai/metrics/surface_dice.py

@@ -282,7 +282,7 @@ def compute_surface_dice(
             spacing=spacing_list[b],
             use_subvoxels=use_subvoxels,
             symmetric=True,
-            class_index=c,
+            class_index=absolute_c,
             warn_empty=warn_empty,
         )
         boundary_correct: int | torch.Tensor | float

monai/metrics/utils.py

@@ -106,6 +106,11 @@ def create_ignore_mask(y: torch.Tensor, ignore_index: int | None) -> torch.Tenso
             values (e.g., -1): they are only meaningful class labels for label-encoded targets. For
             one-hot inputs, negative values fall through to the sentinel path (masking all-zero pixels).
 
+            For one-hot inputs with a valid class index, the mask zeroes all channels at pixels
+            where the ignored class is present (spatial masking). This differs from per-channel
+            exclusion — every pixel belonging to the ignored class is excluded from ALL class scores,
+            not just the score for that class.
+
     Returns:
         Mask tensor of shape (B, 1, H, W, [D]) where 1=valid, 0=ignore.
         Returns None if ignore_index is None.
@@ -460,6 +465,11 @@ def get_edge_surface_distance(
             areas = (areas[0], areas[0])
         elif len(areas) != 2:
             # Unexpected length, create empty tensors
+            warnings.warn(
+                f"Unexpected number of area tensors from get_mask_edges: {len(areas)}, expected 2. "
+                "Falling back to empty tensors.",
+                stacklevel=2,
+            )
             areas = (torch.tensor([], device=y_pred.device), torch.tensor([], device=y_pred.device))
 
     out = convert_to_tensor(((edges_pred, edges_gt), distances, tuple(areas)), device=y_pred.device)  # type: ignore[no-any-return]