Fix: robust channel-last detection for y using num_classes (refs #8366)

Author: 林旻佑

monai/metrics/meandice.py

@@ -310,8 +310,11 @@ def __call__(self, y_pred: torch.Tensor, y: torch.Tensor) -> torch.Tensor | tupl
             y: ground truth with shape (batch_size, num_classes or 1, spatial_dims...).
         """
         y_pred, _ = ensure_channel_first(y_pred)
-        if y.ndim == y_pred.ndim and (y.shape[-1] == y_pred.shape[1] or y.shape[-1] == 1):
+
+        n_ch = self.num_classes or y_pred.shape[1]
+        if y.ndim == y_pred.ndim and y.shape[-1] in (1, n_ch):
             y, _ = ensure_channel_first(y)
+        
 
         _apply_argmax, _threshold = self.apply_argmax, self.threshold
         if self.num_classes is None: