Add AutoAdjustChannel transform

[Nic-Ma]

Is your feature request related to a problem? Please describe.
Now we have AddChannel and AsChannelFirst transforms, need users to manually select expected transform to convert their data to be "channel-first". Several internal researchers got confused by the wrong transform, especially when their image and label have different shapes(maybe image is 3 channels data and channel last, label is single channel data and doesn't have channel dim).
Actually, we can automatically adjust channel dim for most cases based on the spatial_size from meta_data. If advanced users really want to do some special logic, they can use AddChannel and AsChannelFirst directly.

[rijobro]

So always assume that the channel dimension is smaller than all the other dimensions? That seems like a reasonable assumption most of the time!

[Nic-Ma]

Hi @rijobro ,

Thanks for your comments.
I submitted a PR #1675 to address this ticket, will finish it soon.
Actually, all the existing image_readers expect the image to be channel_last or no channel, so I just added a meta info and then automatically adjust the channel dim based on it.

Thanks.

[rijobro]

Ok, do you think the inverse is trivial? I guess it would only require knowing which dimension was the original channel.

[Nic-Ma]

Hi @rijobro ,

Yes, because all the data in MONAI should be channel_first and you can get the original_channel_dim from meta dict. Then I think you can restore the original channel dim or remove channel dim, right?

Thanks.

[rijobro]

Perfect, thanks!