Splenic volume approximation using Deep Learning

[Meddebma]

Hi,
I used the spleen segmentation algorithm to segment the decathlon and my own dataset, and wanted to know if I could integrate a function for measuring the splenic volume (in mm3 or mL).
Thanks!

[Nic-Ma]

Hi @rijobro and @wyli ,

Could you please help share some comments?

Thanks.

[rijobro]

Hi, if you're only doing binary segmentation then you can simply count the number of voxels greater than 0. You could add a new function to your post-transforms, something along the lines of (completely untested):

class GetVol(MapTransform):
    def __init__(keys, select_fn = lambda x: (x>0).sum(), output_key_suffix="_vol"):
        super().__init__(keys)
        self.select_fn = select_fn
        self.output_key_suffix = output_key_suffix
    def __call__(data):
        d = dict(data)
        for key in self.key_iterator(d):
            vol = self.select_fn(d[key])
            d[key + self.output_key_suffix] = vol
        return d

If your data is a simple tensor with batch (i.e., [B,C,H,W,[D]], then you could just do (again, untested):

vols = [(b>0).sum() for b in inferred_seg_batch]

Once you have the number of voxels, you simply multiply on the spacing per voxel to get the volume.

[rijobro]

(1.5, 1.5, 2.0) is the correct spacing to use. For calculating the number of voxels in the segmentation, that'll depend on post_trans. val_output needs to be a binary mask, so you'll need to check that's the case.

[Meddebma]

Yes in fact it is binary I tried the formula vols = [(b>1).sum() for b in val_outputs] and I've got 0, so every point should be a 0 or 1. That is why I dont get it why its making the voxels of the image not of the segmentation mask

here is my code, it is the last part evaluation with other dataset:
https://github.com/Meddebma/pyradiomics/blob/master/Spleen0221.ipynb

[Meddebma]

Hi @rijobro,

should I mabye add AsDiscrete Transform to postprocessing? would that helps?

Thank you so much

[rijobro]

I just tried running your notebook in google colab but it hit some errors. if you'd like me to run it, could you make sure that it works in colab?

[Meddebma]

Hi @rijobro,

I changed the formula to get the pixels of the mask:
vols = [(b[:,1]==1).sum() for b in val_outputs]
it worked :)

thank you very much!