How to handle RGB 2D images

[project-monai]

How to handle RGB 2D images

[ericspod]

TIFF files are often an issue since it's just a broad definition. Typically we want to get 3D images from photographi files with shape (width, height, channel) where channel is 3 for RGB or 4 for RGBA. There are more exotic formats that will have 2 channels for alpha+luminance or other things, but TIFF and others might have a shape (width, height) if they are just greyscale.

One solution to this is to have a transform which adds a 1-size channel if the image is 2D. With this regularized you now need to transpose the images so that they are in (channel, height, width) or (channel, width, height) format to match the Pytorch expectation of channel first, that's why the spatial tranforms misbehave as they are treating the channel dimension as a spatial one.

This sequence of transforms will do this for a sequence of images, the dictionary versions are present in MONAI to do the same with your dictionary data:

test_transforms = Compose(
    [
        LoadImage(image_only=True, reader=PILReader()),
        Lambda(lambda im: im[...,None] if im.ndim==2 else im),
        AsChannelsFirst(),
        Resize(crop_size, "area"),
        ToTensor(),
    ]
)

[mangotee]

Yes, adding AsChannelFirstd(keys=["img", "seg"]) to the transforms list did the trick.
Thanks also for the info regarding TIFFs. It's an image format that I come across rarely. I realized that LoadImaged (ITK image loader) can handle TIFs nicely, and I wrote a class that converts it numpy and makes sure that there are 3 color channels... but in the end I just decided to convert those few TIFFs in my training set to PNG, since I don't expect them to come up during deployment. Thanks anyway for the quick help!