Reslicing a 3D data at any arbitrary angle

[vikashg]

Is your feature request related to a problem? Please describe.
Is there a method to reslice a 3D volume along any random plane. This will be really useful for data-augmentation tasks.
Describe the solution you'd like
A transform that takes a 3D volume and direction cosines for a plane along which the data needs to be resliced. It should also take the slice thickness for the reslicing operation

Describe alternatives you've considered
I have considered using the VTK view port functionality to reslice. But not successful yet.

Additional context
@AndreasHeumann worked on Clara-Viz which is now working with monai-deploy. I wonder if it uses a VTK backend for the rendering. It will be great, if he can share some insight on this.

[Nic-Ma]

Hi @ericspod @AndreasHeumann ,

Do you have any idea about this question?

Thanks in advance.

[AndreasHeumann]

No, Clara Viz is not using a VTK backend for the rendering, it's using our own CUDA based renderer. The renderer includes code to re-orient a 3D volume (swap and flip axes) but not to on a random plane.

[wyli]

I had a quick look and it is supported by the RandAffine transform

import matplotlib.pyplot as plt
from monai.transforms import RandAffine, LoadImage

image = LoadImage(image_only=True)("/Users/Documents/MarsAtlas-MNI-Colin27/colin27_MNI.nii")
print(image.shape)  # (181, 217, 181)
xform = RandAffine(prob=1, rotate_range=(1, 1, 1), spatial_size=(110, 120, 1))

slice_2d = xform(image[None])
print(slice_2d.shape)  # (1, 110, 120, 1) single-channel, (110, 120, 1) volume
plt.imshow(slice_2d[0, :, :, 0])
plt.show()

the idea is to set the output spatial size to (H, W, 1). The slice thickness is essentially about setting scale_range, note that if you specify scale_range=0.2, the actual thickness will be a random value from [1.0-0.2, 1.0+0.2] according to this

MONAI/monai/transforms/spatial/array.py

Line 1433 in bcbfbdd

self.scale_params = self._get_rand_param(self.scale_range, 1.0)

if the input is a torch tensor on cuda device, this computation will be using the gpu.

[wyli]

you can also use monai.transforms.Affine(affine=..., spatial_size=(H, W, 1)), in this case the input affine should be generated beforehand, it can be a transform that combines random aug factors, direction cosines, and thickness.

[vikashg]

Thanks @wyli and @Nic-Ma I will try this today.

Vikash

[vikashg]

Hi @wyli and @Nic-Ma, I tried RandAffine transform and it works great if we want to extract the individual slices as you mentioned. But if I have to rotate the whole 3D volume, I feel that can be done by changing the spatial size to something like (110, 120, 120). I tried doing it and the transform works but the image is displaced as shown here. I think if I can specify the center of rotation in the affine transform it will really help to keep the image in center. How can I specify the center of rotation ? Any thoughts ? Thanks

[wyli]

That's great! for rotate_range translate_range in RandAffine, they are random factors relative to the image center. the goal of this transform is mainly to combine different spatial transforms and only do the resampling once. if you are only interested in rotation, there's a RandRotate

MONAI/monai/transforms/spatial/array.py

Line 945 in c3db337

class RandRotate(RandomizableTransform):

(I'm converting this to a discussion as it's not about the original topic now.)

[vikashg]

I tried this and as you said that the images are rotated based on the center of the image. If I have to rotate the image around a given point, Can I just translate that point in the image to the center of the image and then translate. I was wondering if there is a better way to achieve the same goal.

[vikashg]

Thanks @wyli