LoadImage support for TIFF files

Is your feature request related to a problem? Please describe. Currently due to the extra consideration needed for TIFF files of histopathology WSI images in Camelyon dataset (huge images), we were using PatchWSIData to load these images. However, with other datasets like PANDA dataset for prostate cancer, the size of images allow for loading the entire image into memory. Thus, it might be beneficial to load such images by just using LoadImage transform similar to radiology use cases.

[cgohlke]

tifffile uses multithreads (# of cpus) internally while reading whole image(but it doesn’t support partial image reading)

Tifffile provides zarr store interfaces, which support partial image reading through zarr or dask. Unfortunately the overhead by zarr can be detrimental depending on how the image data in the TIFF is chunked. Tifffile also allows to iterate over segments and to memory map contiguous image data to numpy arrays if possible (e.g. ImageJ hyperstacks).

If multithreads are used in cuCIM, cuCIM is faster than tifffile.

It’s not unusual that C++/GPU implementations outperform Python implementations by 10 to 1000 and scale almost infinitely better with threads. It’s more surprising that tifffile is often in the same order of magnitude when compared to libtiff, Bio-Formats etc. Anyway, the main purpose of tifffile is to get access (from Python) to image and metadata in TIFF-like (non-conformant) files used in bioimaging.

[drbeh]

@gigony, I confirm that tifffile is 3X faster than cuCIM. Is it something expected? That shapes are not the same but it seems that tifffile is not padding.

t0 = perf_counter()
img_cucim = CuImage(filename).read_region(level=1)
t_cucim = perf_counter() - t0
print(f"CuCIM   : shape = {img_cucim.shape}, time = {t_cucim:.3f}")

t0 = perf_counter()
img_tiff = tifffile.imread(filename, key=1)
t_tiff = perf_counter() - t0
print(f"Tifffile: shape = {img_tiff.shape}, time = {t_tiff:.3f}")

CuCIM   : shape = [110592, 48896, 3], time = 29.703
Tifffile: shape = (110226, 48734, 3), time =  8.446