Add CropOrPadAtCenter transform class

[themantalope]

🚀 Feature A crop or pad function that allows the user to crop/pad an image but specify where the center of the new image should be based on the input image. For example, let's say I have an organ centered at location [45,62,101] in a volume that is [368, 512, 128] in size and I want my new image to be [128,128,128] in size, and the center of the new image (i.e. [64,64,64]) to map to the point [45,62,101] in the original image.

Motivation

Developing datasets with tumors, need to crop and pad around the tumors.

Pitch

Described above. Proposed code below.

Alternatives

N/A

Additional context Here is my proposed code. I've tested it locally but probably need more robust testing.

import torchio as tio

class CropOrPadAtCenter(tio.Transform):
    
    def __init__(self, center , target_shape, image_or_mask_name=None, **kwargs):
        super().__init__(**kwargs)
        self.target_shape = target_shape
        self.center = center
        self.image_or_mask_name = image_or_mask_name
        

    def apply_transform(self, subject):
        # if image or mask name is None, use the first image as the base image to work with
        if self.image_or_mask_name is None:
            images = list(subject.keys())
            base_image = images[0]
            self.image_or_mask_name = base_image
        
        image = subject[self.image_or_mask_name]
        non_channel_im_shape = image.shape[-3:]
        
        # first assert that the center is in the image
        assert all([c >= 0 and c < s for c, s in zip(self.center, image.shape[-3:])])
        # next determine if we need to pad. if the bounds of the target shape are outside the image, we need to pad
        
        # compute how many pixels we need to pad for each dimension 
        pad = []
        for c, s, t in zip(self.center, non_channel_im_shape, self.target_shape):
            
            lower = 0
            upper = 0
            
            if c - t//2 < 0:
                lower = abs(c - t//2)
            
            if c + t//2 > s:
                upper = c + t//2 - s

            pad.extend([lower, upper])
        # pad the image
        p = tuple(pad)
        pad_xform = tio.Pad(p)
        subject = pad_xform(subject)

        # now crop the image
        # the crop function expects the start and dim_size - end of the crop (weird)
        image = subject[self.image_or_mask_name]
        non_channel_im_shape = image.shape[-3:]
        lower_bound_pads = [p for i, p in enumerate(pad) if i % 2 == 0]
        new_center = [c + l for c, l in zip(self.center, lower_bound_pads)]
        self.center = new_center
        crop = []
        width = self.target_shape[0]//2
        height = self.target_shape[1]//2
        depth = self.target_shape[2]//2
        im_width, im_height, im_depth = non_channel_im_shape
        for d, s, c in zip([width, height, depth], [im_width, im_height, im_depth], [self.center[0], self.center[1], self.center[2]]):
            start = c - d
            end = s - (c+d)
            crop.extend([start, end])

        ct = tuple(crop)
        crop_xform = tio.Crop(ct)
        subject = crop_xform(subject)
        return subject

        

[romainVala]

Hi thanks for sharing. Actually the same behavior can be achieve with torchio LabelSampler

here is a code exemple:

import torchio as tio, torch

center_patch = [132,100,170]
patch_size = 128
around_center = 0

suj = tio.datasets.Colin27()
label_proba =  torch.zeros_like(suj.t1.data)
label_proba[0,center_patch[0]-around_center:center_patch[0]+around_center+1,center_patch[1]-around_center:center_patch[1]+around_center+1,center_patch[2]-around_center:center_patch[2]+around_center +1] = 1

img = tio.LabelMap(tensor = label_proba, affine=suj.t1.affine)

suj.add_image(img, 'label_proba')

#Pad to handel worste case cenario where center is at the border of the image (so extend with patch_size/2 + around_center//2 )

t_pad = tio.Pad(patch_size//2 + around_center//2)
suj = t_pad(suj)

lab_s = tio.LabelSampler(128,'label_proba',{0:0,1:1})
generator = lab_s(suj, num_patches=10)

for patch in generator:
    locations = patch[tio.LOCATION]
    print(f'loc {locations} ')
#note here that locations are 3 indices start and 3 indice end in the padded suj indexes (not the center patch)
# so it is different from the original center_patch because we pad the volume, but the results is as expected

Does will suite your needs ?

[themantalope]

I see, yes that would work for what I'm looking for. May need some additional code to ensure that I'm getting all voxels in the sampled image ,but this is basically what I want.

[romainVala]

what do you mean by all voxel ? You will get only voxel that are less than patch_size/2 away from the center

[themantalope]

to ensure that the sampled patch contains all positive voxels in the label.

i guess the difference here is that i'm specifying the center so there is more control over where the sample is getting drawn from, which is important for my use case