Lazy loading of numpy arrays

[malteekj]

Is it possible to do lazy loading of numpy arrays by passing a path to the ScalarImage and LabelMap class?
I have my data as numpy arrays and can't have it all in memory.
e.g.:

import torchio as tio
subj = tio.Subject(
    CT = tio.ScalarImage(path = path_to_numpy_CT),
    label = tio.LabelMap(tensor = path_to_numpy_ROI)
)

I keep getting the error message "RuntimeError: "path_to_numpy_file" not understood.

[fepegar]

Hi, @malteekj.

I keep getting the error message "RuntimeError: "path_to_numpy_file" not understood.

This is because NumPy is not a suitable format for medical images, as it doesn't store important spatial metadata. Therefore, it's not supported by SimpleITK and NiBabel, which TorchIO uses for I/O.

I've added a kwarg in Image so you can use your own reader for lazy loading:

import numpy as np
import torchio as tio

def numpy_reader(path):
    return np.load(path), np.eye(4)

subj = tio.Subject(
    ct=tio.ScalarImage(path_to_numpy_CT, reader=numpy_reader),
    label=tio.LabelMap(path_to_numpy_ROI, reader=numpy_reader)
)

I will merge these changes soon, they will be available in v0.18.27.

I hope that works for you.

[malteekj]

Thanks! This is was I was looking for.

And thanks for the great work you have done with this repo.