Forcing BCE for DiceCELoss when overlapping labels predictions

[SebGoll]

Hello,

I have encountered multiple projects in which I need to output multiple labels for a single voxel. In this case, I'd like to use the DiceCELoss, but forcing it to use the binary cross entropy loss instead of the cross entropy one (since each channel predicted is either on or off independently of the other channels).
In the current state of DiceCELoss, it automatically choose to use either ce/bce based on the number of channels:

MONAI/monai/losses/dice.py

Line 804 in e72145c

ce_loss = self.ce(input, target) if input.shape[1] != 1 else self.bce(input, target)

I wanted to see if there is an officially supported way to achieve the behavior described above before contributing.

Thank you

[denis-samatov]

Yes — for overlapping labels, each output channel is an independent binary target, so the CE branch of DiceCELoss is not the loss you want. In MONAI 1.6.0, DiceCELoss.forward chooses CrossEntropyLoss whenever the prediction has more than one channel; sigmoid=True changes only its Dice component. There is no force_bce switch for multi-channel input.

You can compose the two supported losses explicitly:

import torch
from monai.losses import DiceLoss

dice = DiceLoss(
    sigmoid=True,
    softmax=False,
    to_onehot_y=False,
    include_background=True,
)
bce = torch.nn.BCEWithLogitsLoss()

def multilabel_loss(logits: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
    # Both tensors: [B, C, *spatial]; target is multi-hot (0/1), overlaps allowed.
    target = target.to(dtype=logits.dtype)
    return dice(logits, target) + bce(logits, target)

Pass raw logits to both terms: DiceLoss applies its own sigmoid, and BCEWithLogitsLoss incorporates the sigmoid numerically stably. Do not use softmax, argmax, or to_onehot_y=True for independent overlapping labels. If all your channels are foreground labels, keep include_background=True so channel 0 is not dropped; change it only if channel 0 really represents a separate background channel. For inference, use torch.sigmoid(logits) and threshold each channel independently.

I checked this with a synthetic voxel belonging to two labels on MONAI 1.6.0 / PyTorch 2.9.0: the composed loss was finite and gradients propagated. That verifies the API path, not performance on a real segmentation dataset. If you need imbalance handling, add pos_weight to BCEWithLogitsLoss based on your training data rather than assuming the default weighting is optimal.