DiceMetric: mean_batch and mean_channel reduction modes appear to be swapped #8602
Description
Describe the bug
The mean_batch and mean_channel reduction modes in DiceMetric.aggregate() appear to be swapped. mean_batch returns one value per channel (instead of per batch), and mean_channel returns one value per batch (instead of per channel).
Is this the intended behavior, or am I misunderstanding how data should be accumulated or passed to DiceMetric? In the below example, the dimension of the data is BCHW.
To Reproduce
Run:
import torch
from monai.metrics import DiceMetric
n_batches = 3
batch_size, n_classes, h, w = 4, 5, 128, 128
y_pred_batches = [torch.rand(batch_size, n_classes, h, w) for _ in range(n_batches)]
y_batches = [torch.rand(batch_size, n_classes, h, w) for _ in range(n_batches)]
dm = DiceMetric()
for i in range(n_batches):
dm(y_pred_batches[i], y_batches[i])
# Internal buffer has shape [12, 5] (12 samples total, 5 channels)
print(f"Buffer shape: {dm.get_buffer().shape}") # [12, 5]
# mean_batch returns 5 values (one per channel, not per batch)
print(f"mean_batch: {dm.aggregate(reduction='mean_batch').shape}") # [5]
# mean_channel returns 12 values (one per batch, not per channel)
print(f"mean_channel: {dm.aggregate(reduction='mean_channel').shape}") # [12]Expected behavior
mean_batch should average across channels and return one value per batch: shape [3]
mean_channel should average across batches and return one value per channel: shape [5]
Environment
Ensuring you use the relevant python executable, please paste the output of:
MONAI version: 1.5.0
Numpy version: 2.2.6
Pytorch version: 2.6.0+cu124
Additional context
The problem appears to be in monai/metrics/utils.py:do_metric_reduction() lines 113-121. I am not sure if the bug affects other metrics using this function too.