Crop a fixed-size region surrounding tumor centroid

[mattwarkentin]

Hi,

I am trying to perform a transform that, for each subject, will crop a volume of fixed size centred around a coordinate. The coordinate corresponds to the center of the tumor in 'mask'. This is a pretty common task. However, the coordinate varies for each subject because it corresponds to the location of a tumor. It seems like when you instantiate the SpatialCrop/SpatialCropd transforms you need to input a roi_center, but shouldn't this support centers that are patient-specific?

Something like this seems like it would make sense:

sub = {
  'img': 'image.nrrd',
  'mask': 'mask.nrrd',
  'coord': (100, 110, 75)
}

SpatialCropd(['image'], roi_center_key='coord', roi_size=(50, 50, 50))

The code below is as close as I've gotten to what I desire so far, but it firsts crops based on the masks bounding box, then resizes to the desired dimension (50, 50, 50). But if the mask is smaller than 50x50x50, the final image centers on the immediate nodule bounding box and is padded with black space, not ideal.

keys = ['img', 'mask']

trans = Compose([
  LoadImaged(keys),
  AddChanneld(keys),
  CropForegroundd(keys, source_key='mask'),
  ResizeWithPadOrCropd(keys, spatial_size=(50, 50, 50)),
  ToTensord(keys)
])

sub_t = trans(sub)

What I want is a crop of 'img' that is 50x50x50 in size and is centred on the tumor.

[rijobro]

Hi @mattwarkentin , I think the easiest thing would be to create your own transformation that inherits from MapTransform. During the __call__, you would extract the coordinates from one of the other keys in your data.

Something like this (untested):

class DynamicCrop(MapTransform):
    def __init__(self, keys, coord_key):
        super().__init__(keys)
        self.coord_key = coord_key
    def __call__(self, data):
        d = dict(data)
        coord = d[self.coord_key]
        cropper = SpatialCrop(...)  # use the coordinates and your desired size to create a cropper
        for k in self.keys:
            d[k] = cropper(d[k])

[mattwarkentin]

Great suggestion! Much appreciated. Still learning the flexibility of MONAI.

[mattwarkentin]

For posterity, this function works as described in the original post and will ALWAYS return an object of the desired size...

class DynamicCropOrPad(monai.transforms.MapTransform):
    def __init__(self, keys, coord_key, roi_size):
        super().__init__(keys)
        self.coord_key = coord_key
        self.roi_size = roi_size
    def __call__(self, data):
        d = dict(data)
        coord = d[self.coord_key]
        cropper = SpatialCrop(roi_center=coord, roi_size=self.roi_size)
        resizer = ResizeWithPadOrCrop(spatial_size=self.roi_size)
        for k in self.keys:
            d[k] = cropper(d[k])
            if (tuple(d[k].shape) != self.roi_size):
              d[k] = resizer(d[k])
        return(d)

[rijobro]

Great, glad you got it working!

[LalehSeyyed]

Hi @mattwarkentin, I am wondering how do you define coord_key. as you mentioned the coord_key varies per patient based on the location of the tumor. Do you feed the location of the tumor from your dataset or do you use a code to automatically locate the coord_key from your segmentation mask labels?

[rijobro]

Although the tumour centre varies between patients, the key used to store that information in the data dictionary can remain constant.

For example, your data might have the following structure:

data = {
    "image": torch.Tensor(...),
    "label": torch.Tensor(...),
    "tumor_center": torch.Tensor(...),  # coordinates that point to tumour centre. can change between patients, but key is always the same.
}

Edit - sorry I completely misread your question @LalehSeyyed! There are lots of way of generating the coordinates, using another network, for example. I'll let @mattwarkentin let you know how he does it in his specific case.

[mattwarkentin]

Hi @LalehSeyyed. In my case, the coordinates for the centre of the tumour were reported by radiologists who had previously reviewed the scans. But as @rijobro says, the coordinate value could be obtained using any method really. If you have a segmentation mask, you could also probably obtain coordinates by finding the centre of a label but it might take some extra code to get that information.