Improving the docstring explanation

Signed-off-by: Vijay Vignesh Prasad Rao <vijayvigneshp02@gmail.com>

Author: VijayVignesh1

monai/metrics/meandice.py

@@ -13,6 +13,7 @@
 
 import torch
 
+from monai.apps.detection import metrics
 from monai.metrics.utils import compute_voronoi_regions_fast, do_metric_reduction
 from monai.utils import MetricReduction, deprecated_arg
 from monai.utils.module import optional_import
@@ -47,8 +48,17 @@ class DiceMetric(CumulativeIterationMetric):
     image size they can get overwhelmed by the signal from the background. This assumes the shape of both prediction
     and ground truth is BCHW[D].
 
-    The ``per_component`` parameter can be set to `True` to compute the Dice metric per connected component in the ground truth
-    , and then average. This requires binary segmentations with 2 channels (background + foreground) as input.
+    The `per_component=True` approach computes the Dice metric on a per-connected component basis in the ground truth segmentation,
+    ensuring equal weighting for each component regardless of its size. This method eliminates biases in traditional metrics,
+    providing a more balanced evaluation, particularly in scenarios where object size does not correlate with clinical relevance.
+    This provides a more granular evaluation of segmentation quality, especially useful when dealing with fragmented or
+    disconnected objects in the foreground.
+    Note:
+    - The input prediction (`y_pred`) and ground truth (`y`) must both have 2 channels (foreground/background),
+    with binary segmentation (0 for background, 1 for foreground). That is, this assumes the shape of both prediction
+    and ground truth is B2HW[D].
+    - This method cannot be used with multiclass segmentation.
+    For more information, refer to the original paper: https://arxiv.org/abs/2410.18684
 
     The typical execution steps of this metric class follows :py:class:`monai.metrics.metric.Cumulative`.
 

tests/metrics/test_compute_meandice.py

@@ -346,17 +346,11 @@ def test_nans_class(self, params, input_data, expected_value):
     @unittest.skipUnless(has_ndimage, "Requires scipy.ndimage.")
     def test_cc_dice_value_nogpu(self, params, input_data, expected_value):
         dice_metric = DiceMetric(**params)
-        cpu_inputs = {"y": input_data["y"].cpu(), "y_pred": input_data["y_pred"].cpu()}
-        dice_metric(**cpu_inputs)
-        result = dice_metric.aggregate(reduction="none")
-        np.testing.assert_allclose(result.cpu().numpy(), expected_value, atol=1e-4)
-
-    @parameterized.expand([TEST_CASE_16, TEST_CASE_17])
-    @unittest.skipUnless(has_ndimage, "Requires scipy.ndimage.")
-    @unittest.skipUnless(torch.cuda.is_available() and has_cupy_ndimage, "Requires CUDA and cupyx.scipy.ndimage.")
-    def test_cc_dice_value_gpu(self, params, input_data, expected_value):
-        dice_metric = DiceMetric(**params)
-        dice_metric(**input_data)
+        if not has_cupy_ndimage:
+            cpu_inputs = {"y": input_data["y"].cpu(), "y_pred": input_data["y_pred"].cpu()}
+            dice_metric(**cpu_inputs)
+        else:
+            dice_metric(**input_data)
         result = dice_metric.aggregate(reduction="none")
         np.testing.assert_allclose(result.cpu().numpy(), expected_value, atol=1e-4)