Question: slow torchio.ScalarImage compared to pydicom.dcmread

[vedal]

Versions:

python 3.8.5
torch==1.13.0+cu116
torchio==0.18.86
pydicom==2.3.1

Hi!
Thanks for this great package!
I find that loading a dicom using pydicom.dcmread is much faster than torchio.ScalarImage. I tried avoiding the NaN-check. Is there any other method to speed it up?

%timeit -n 5 torch.tensor(pydicom.dcmread(img_path).pixel_array)
>>> 658 ms ± 9.19 ms per loop (mean ± std. dev. of 7 runs, 5 loops each)

%timeit -n 5 torchio.ScalarImage(img_path, check_nans=False).data
>>> 2.39 s ± 28.4 ms per loop (mean ± std. dev. of 7 runs, 5 loops each)

The dicom data has shape (80, 758, 1016, 3), which torchio.ScalarImage automatically reshapes to torch.Size([3, 1016, 758, 80]).

Thanks in advance :)

[fepegar]

Hi, @vedal. Thanks for your kind words and for sharing this. We could consider switching to pydicom for loading DICOM. Can you please share some data to reproduce your results? This is what I'm getting:

In [15]: %timeit torch.tensor(pydicom.dcmread('IMG0001.dcm').pixel_array)
1.94 ms ± 728 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

In [16]: %timeit torch.tensor(tio.ScalarImage('IMG0001.dcm').data)
<magic-timeit>:1: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
1.65 ms ± 484 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

[vedal]

Thanks for your response @fepegar.
Hm, it might depend on size of the dicom then. I'm not sure how large or how many frames your dicom was. Maybe its a thing of multi-frame dicoms? For example, the following multi-frame dicom I found online, shape=((42, 576, 768, 3) seems to already show the contours of the problem:

In [8]: %timeit -n 10 torch.tensor(pydicom.dcmread('ileitis_in_yersiniosis_42_0001.dic').pixel_array)
>>> 207 ms ± 1.59 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [9]: %timeit -n 10 tio.ScalarImage('ileitis_in_yersiniosis_42_0001.dic').data
>>> 691 ms ± 1.29 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

Note that I only convert to torch.tensor in pydicom for fairness, since tio.ScalarImage already returns a tensor.

[fepegar]

Given that TorchIO uses SimpleITK to read DICOM, I suggest you open an issue in their repo.

[vedal]

@fepegar is it an option to change to pydicom under-the-hood in torchio, as you suggested in your previous answer?

[fepegar]

I suppose you can define a function that takes a path, reads the data with pydicom and returns a 4D tensor and a 2D affine, as specified in https://torchio.readthedocs.io/data/image.html#torchio.Image