Why isn't 'include_background' named 'include_first_channel'?

[ellisdg]

In many of the loss classes and metrics in MONAI, the creators have put in an argument named include_background. To my understanding, what this argument does is include the first channel in the computations, if True, and exclude the first channel in the computations, if False.

I have a few issues with the naming of the include_background argument:

1. The name is ambiguous and does not reflect what the argument actually does.
2. The background channel could be any of the channels, and does not necessarily have to be the first channel.
3. When there is no background channel, you should specify include_background=True, which is counter-intuitive because there is no background to be included.

Therefore, why isn't the argument called include_first_channel instead? This name says what the include_background argument actually does.

[KumoLiu]

Hi @ellisdg, thanks for your interest and suggestion here.
We used 'include_background' to specify whether we calculated with or without the first category, which is usually assumed to be background. Using 'include_first_channel' would be misleading because what we want to do is determine whether or not to include background. Does it make sense?

MONAI/monai/metrics/meandice.py

Lines 30 to 32 in a4e4894

	`y_pred` is expected to have binarized predictions and `y` can be single-channel class indices or in the
	one-hot format. The `include_background` parameter can be set to ``False`` to exclude
	the first category (channel index 0) which is by convention assumed to be background. If the non-background

Additionally, when only single-channel class indices exist, we can also specify whether to include background.

MONAI/monai/metrics/meandice.py

Lines 253 to 258 in a4e4894

	first_ch = 0 if self.include_background else 1
	data = []
	for b in range(y_pred.shape[0]):
	c_list = []
	for c in range(first_ch, n_pred_ch) if n_pred_ch > 1 else [1]:
	x_pred = (y_pred[b, 0] == c) if (y_pred.shape[1] == 1) else y_pred[b, c].bool()

I hope it helps, thanks!

[ebrahimebrahim]

As a MONAI user when I saw include_background, I had to check at the documentation to make sure that by "background" it meant the zeroth channel. include_first_channel or include_zeroth_channel would have been clearer to me.

[ellisdg]

@KumoLiu, thank you for your reply!

As a MONAI user when I saw include_background, I had to check at the documentation to make sure that by "background" it meant the zeroth channel. include_first_channel or include_zeroth_channel would have been clearer to me.

I agree with @ebrahimebrahim. Either of those suggestions would increase the clarity and ultimately reduce the amount of errors that users make. include_channel_zero could be another option.

Using 'include_first_channel' would be misleading because what we want to do is determine whether or not to include background.

The include_background flag is misleading when you have a single or multi-channel output and there is no background channel, because setting include_background=True does not actually include the background.

Additionally, when only single-channel class indices exist, we can also specify whether to include background.

The example code you provided reinforces my point that when we say include_background we are talking about the first channel (or zeroth_channel as @ebrahimebrahim suggested). When include_background is True, then first_ch is set to 0, which forces the first channel to be included. When include_background is False, then first_ch is set to 1, which forces the first channel to be excluded. I think getting rid of the include_background flag altogether and letting the user specify the first channel would be more intuitive than the include_background flag. It doesn't matter if we are talking about actual channels or label indices, the code you provided uses the first_ch variable for both.

In my opinion, the first_ch variable would be better named something along the lines of first_evaluated_channel, first_eval_ch, or start_channel, because the first channel doesn't ever change from being channel zero, but the first channel being evaluated can change.

[KumoLiu]

Thanks, @ellisdg and @ebrahimebrahim for your suggestions.
If that is the general confusion among MONAI users, I believe we should reconsider the name. Let's bring it up to more people here. cc @wyli @ericspod @Nic-Ma @vikashg

[ericspod]

It has been convention that background was always channel 0 and this has made sense for the segmentation that we've been doing. I can see other applications where there is no background as described here so changing the name probably should be done, eg. "include_channel_zero" or something like that. This variable name is one of those, shall we say, "historical" parts of MONAI. I would still be explicit in documentation and elsewhere that the background channel is channel 0 if any.

[ellisdg]

This is might be a naive question, but what is the benefit to having a background channel if it isn't included in the loss function?

[ericspod]

If you have a multi-class segmentation problem where the segments are relatively small sometimes the loss function won't penalise the network enough if it predicts all background. The background signal in this case tends to dominate and the network can forget about smaller classes. You still need a background channel of course because every pixel/voxel must have a prediction. For example, cardiac MR segmentation of the ventricles is where I've encountered this behaviour.

[ellisdg]

Thank you for your response!

You still need a background channel of course because every pixel/voxel must have a prediction.

You can also use a sigmoid activation function and then assign the background to any voxels that don't pass the activation threshold. That is the strategy taken in this MONAI tutorial for BRATS tumor segmentation:
https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/brats_segmentation_3d.ipynb

In this case the label channels are tumor core, whole tumor, and enhancing tumor, and there is no background channel. So I guess I am still confused as to why the default behavior should be that all models have a background channel.

[ericspod]

Using sigmoid is something we've done with binary segmentation where the prediction tensor has only 1 channel. For doing this with multi-class I guess it's possible but not what I've seen done and may not perform well with Dice or other losses.

[ChenglongWang]

I believe that using "include_channel_zero" as an alternative to "include_background" may not be the best choice due to the loss of argument semantics. In contrast, the name "include_background" directly conveys the purpose of this argument to the user. The core issue of original post is that the channel of background is confusing. One potential solution could be to introduce an optional argument, such as background_channel: int | Sequence[int] = 0, to clarify this aspect. BTW, argument such as include_channel: int | Sequence[int] = 0 could be better than "include_channel_zero" .

[ericspod]

Thanks for the feedback, we'll keep this in mind for future revisions.