Support loading raw binary image files

[Nepelius]

🚀 Feature
A loading functionality for .raw image files.

Motivation

I often work directly with .raw files and loading .raw files would be a nice extention of the current dataloading functions using SimpleITK, NiBabel and so on. Many CT-devices for example work primarily with raw files.

Pitch

One possible implementation would be a function read_raw inside src/torchio/data/io.py, where the raw file is opened and read via np.fromfile(). Since raw binary files contain no metadata, the loader would require the image shape, data type and affine (or spacing/orientation information) to be provided explicitly:

• shape: dimensions of the image, for example [456, 400, 512]
• dtype: What data type the image should be loaded in, e.g. np.float32
• affine: Affine matrix
• returns: [tensor, affine]

Alternatives

I am not aware of existing support for loading raw binary data with any imaging framework like MONAI, SimpleITK or NiBabel.

Additional context

I have something similar implemented, this would need some adaptions to integrate to torchio. How the function could look like:

def read_raw(path, shape, dtype, affine):

   # read raw file from filepath
   img = np.fromfile(path, dtype=dtype)

   # reshape 1D array from raw file to given image dimensions
   img = np.reshape(img, (shape[2], shape[1], shape[0]))
   img = np.transpose(img, (2, 1, 0))

   return torch.from_numpy(img), affine

Also: This is my first GitHub issue, I am open for feedback :)