A possible feature: save pixels position in dcmread to enable gpu direct dicom file loading

[yiheng-wang-nv]

Hi pydicom maintainers,

I did some research on pydicom source code, and found that it's feasible to enable GPU Direct Storage to accelerate data loading.
Via using kvikio, we can load a dicom file from disk into GPU directly. If we know the position of pixels, it's easy to parse loaded bytes into a pixel array.

A draft PR is submitted to show the potential changes: #2189
and later I will also post some experiment code on how to make use of the changes and accelerate dicom image loading with GPU.

Welcome to discuss about if it's fine to enable this feature, thanks in advance!

[scaramallion]

Get the offset to the start of the Pixel Data value without loading the actual image data into memory:

>>> from pydicom import examples, dcmread
>>> path = examples.get_path("ct")
>>> ds = dcmread(path, defer_size="100 KB")  # Don't read large element values into memory
>>> ds.get_item(0x7FE00010, keep_deferred=True).value_tell
6300
>>> type(ds._dict[0x7FE00010])
<class 'pydicom.dataelem.RawDataElement'>

[yiheng-wang-nv]

Thanks @scaramallion , it works well.
I can do sth like the following to make use of dcmread and convert bytes in GPU to correct array:

    dcm_read_data = dcmread(file_path, defer_size="100 KB")

    # Extract relevant attributes
    rows = dcm_read_data.Rows
    columns = dcm_read_data.Columns
    bits_allocated = dcm_read_data.BitsAllocated
    samples_per_pixel = dcm_read_data.SamplesPerPixel
    number_of_frames = getattr(dcm_read_data, 'NumberOfFrames', 1)
    pixel_representation = dcm_read_data.PixelRepresentation
    is_little_endian = dcm_read_data.is_little_endian

    if bits_allocated == 8:
        dtype = cp.int8 if pixel_representation == 1 else cp.uint8
    elif bits_allocated == 16:
        dtype = cp.int16 if pixel_representation == 1 else cp.uint16
    elif bits_allocated == 32:
        dtype = cp.int32 if pixel_representation == 1 else cp.uint32
    else:
        raise ValueError("Unsupported BitsAllocated value")

    bytes_per_pixel = bits_allocated // 8
    total_pixels = rows * columns * samples_per_pixel * number_of_frames
    expected_pixel_data_length = total_pixels * bytes_per_pixel

    offset = dcm_read_data.get_item(0x7FE00010, keep_deferred=True).value_tell

    with kvikio.CuFile(file_path, "r") as f:
        buffer = cp.empty(expected_pixel_data_length, dtype=cp.int8)
        f.read(buffer, expected_pixel_data_length, offset)

    cupy_data_array = buffer.view(dtype).reshape((number_of_frames, rows, columns))

I closed the draft PR since there is no need to do that.

[yiheng-wang-nv]

Hi @scaramallion , I think the GDS + pydicom application for faster loading could be a feature, or at least a tutorial contribution for the repository. Do you thinks it's fine to let me contribute it? If so, I will following the contribution guide and submit a formal PR.

Thanks!