Resize a dicom image

[vikashg]

Hello everyone,
I have dicom images which when read using the LoadImage() transform gives me an image of format (1, 1890, 2050). It is a 2D image. The first axis is the channel. However, I would like it to reshape it as an image of size (500, 500).
I was wondering if there was a transform to remove the first channel ?
In addition, when I use the Resize filter as Resize((1, 500, 500)) , it will throw an image of size (1, 1890, 500, 500).
I do not understand this Resize function.
Any help would be really appreciated.

[Nic-Ma]

Hi @vikashg ,

You can use the SqueezeDim transform:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/utility/array.py#L509
And BTW, the MONAI API doc is: https://docs.monai.io/en/stable/api.html

Thanks.

[wyli]

The first parameter to Resize is a 'spatial' size, so in this case it would be (500, 500):

monai.transforms.Resize((500, 500))(torch.zeros((1, 1890, 2050))).shape

https://docs.monai.io/en/stable/transforms.html#monai.transforms.Resize

[vikashg]

Thanks @wyli and @Nic-Ma , Although this works, I also tried something like

monai.transforms.Resize((500, 500))(torch.zeros((1890, 2050, 1))).shape

It yields a shape (1890, 500, 500). The part which is confusing is that the position of channels affect the size of the output.

[vikashg]

In addition,
if I do the following chain of transforms

train_transforms = Compose([LoadImage(image_only=True, reader="itkReader"), SqueezeDim(dim=2),AddChannel(), Resize((500,500)), 
 EnsureType()])

I get an output of the size(1, 1, 500, 500), where the original image was of size (1890, 2050, 1). So I had to do this

train_transforms = Compose([LoadImage(image_only=True, reader="itkReader"), SqueezeDim(dim=2),AddChannel(), Resize((500,500)), SqueezeDim(dim=0), EnsureType()])

to get an image of size (1, 500, 500). I just find the way Resize() is implemented a bit weird. Am I missing something in my implementation?

Thanks once again.

[Nic-Ma]

You can try the EnsureChannelFirst transform or AsChannelFirst instead of current logic.
All the MONAI transforms expect channel-first data shape.

Thanks.

[vikashg]

Ok so I tried this

train_transforms = Compose([LoadImage(image_only=True, reader="itkReader"), AsChannelFirst(), Resize((500,500)),						EnsureType()])

The output is of the shape (1, 1, 500, 500). I feel the extra channel at the end is for the index in the batch of the dataset.

[Nic-Ma]

So is there still any problem in your test now?

Thanks.

[rijobro]

The output is of the shape (1, 1, 500, 500). I feel the extra channel at the end is for the index in the batch of the dataset.

Our transforms don't add a batch dimension (this is only added by the DataLoader). I would suggest confirming the initial shape of your image with:

print(LoadImage(image_only=True)(filename).shape)

[vikashg]

Ok got it, Thanks @Nic-Ma, It is good now so far. I have the output shapes I was looking for. Thanks @rijobro for the pointers.