crop ct scan deep from the top

[oscarramirezs]

HELLO,
I have a question, can someone help me? sorry it may be a silly question

I have some Keen CT scan, each image is 512x512 and deep changes from one CT scan to another (ex; sometime 120 images to 900 images), sometimes the CT start from the ankle and some time from near the knee,

so when I use crop = SpatialCropd(keys=["image", "label"], roi_start=[0, 0, 0], roi_end=[512, 512, 256]), the scanner is cut from the bottom to the top, but I want to start cropping from the top to the Botton because always the keen is in the top.

How can I cut the images starting always from the last one, ? ( the last one change for each ct)

thank you for your help

Oscar

[function2-llx]

I think Orientation is what you want, which reorients images to a fixed orientation. Check the "Reorientation to a designated axes codes" section in this tutorial.

[function2-llx]

Seems that I misunderstood your question at first.

TL;DR: use Lambda(lambda x: x[:, :, -256]).

In fact something like roi_start=(0, 0, -256) and roi_end=(None, None, None) should work, which makes sense like slicing an array, however this currently does not work in MONAI (as I've mentioned in #1935). But some workaround like Lambda(lambda x: x[:, :, -256]) will work perfectly.

[oscarramirezs]

Hi function2-llx

Thank your help,

sorry I'm new to python, where should I use lambda function? I have my transformation function as shown below, if you can help me, I will appreciate it very much.

train_transforms = Compose(
[
LoadImaged(keys=["image", "label"]),
AddChanneld(keys=["image", "label"]),
lambd = Lambda(func=lambda x: x[:, :, -256]),
SpatialCropd(keys=["image", "label"],roi_start=[0, 0, -256], roi_end=[None, None, None]),
RandFlipd(keys=["image", "label"], prob=0.33, spatial_axis=0).set_random_state(rseed),
RandGaussianNoised(keys=["image"], prob=0.1, mean=48, std=330).set_random_state(rseed),
ToTensord(keys=["image", "label"]),
]
)`

thank you
Oscar

[function2-llx]

@oscarramirezs Check if this works:

train_transforms = Compose([
    LoadImaged(keys=["image", "label"]),
    AddChanneld(keys=["image", "label"]),
    Lambdad(keys=["image", "label"], func=lambda x: x[:, :, :, -256]),
    RandFlipd(keys=["image", "label"], prob=0.33, spatial_axis=0).set_random_state(rseed),
    RandGaussianNoised(keys=["image"], prob=0.1, mean=48, std=330).set_random_state(rseed),
    ToTensord(keys=["image", "label"]),
])

[oscarramirezs]

Hi, @function2-llx

it works thanks

I changed it a little for

Lambdad(keys=["image", "label"], func=lambda x: x[ :, :, :, -256:]),

Thanks a lot @function2-llx

Oscar

[function2-llx]

Congrats!

Yeah a typo missing func key, I've already edited it.

[function2-llx]

As the discussion above, we could achieve it with Lambda. However I think it will be cool if I could do the same thing just with SpatialCrop.

[rijobro]

Hi, suggested PR here. If you wanted to leave the last 100 in the z-direction and not touch anything else, this would be done with:

SpatialCrop(roi_slices=[slice(None), slice(None), slice(-100, None)])

[rijobro]

This has the advantage of being invertible.

[function2-llx]

Hi, may I ask that if roi_slices is supported, is it necessary to keep roi_start and roi_end?

[rijobro]

You can use roi_slices or roi_start and roi_end. By the way, the PR has been merged, so you should be able to use it if you're using an up-to-date version of MONAI master.

[rijobro]

I actually just checked the code. If roi_slices is provided, roi_start and roi_end won't be used if they have been provided, but an error won't be thrown either. They will just be ignored.

[Nic-Ma]

Hi @rijobro , I submitted a bug related to this question, could you please help take a look? It's urgent issue for our release: #2087.

Thanks in advance.