Support for find the whole-slide image (WSI) resolution (in microns per pixel, MPP) and extracting tiles with given MPP resolution

[nodp53]

Hi Manuel, at first a real big thx to your for this excellent work. From project design to documentation wau. The first project where everything works as it should. Small suggestion from me would be the implementation of possibility of extraction the tiles with given MPP resolution. Maybe like this:

`import pathlib import re import datetime

import pyvips as vips import pandas as pd from wsi_tile_cleanup import filters, utils

from .timer import Timer

defaults

MPP = 0.5 DIMS = 224 FMT = ".png" OUT_DIR = "."

def rescale_slide(vi, mpp=MPP, mpp_key="openslide.mpp-x"):

TODO: generalize the mpp_key? some alternatives:

# - "aperio.MPP", "openslide.mpp-y", "openslide.mpp-x"
vi_mpp = float(vi.get(mpp_key))
mpp = float(mpp)

scale = vi_mpp / mpp

# 5% "tolerance" because mpp tend to vary within a certain range
if scale > 1.05:
    print(f"- slide will be scaled up by {scale}")
    print("WARNING: beware potential losses")
elif scale < 0.95:
    print(f"- slide will be scaled down by {scale}")
else:
    print(f"- slide will NOT be rescaled | {vi_mpp} is within range of targeted mpp")
    return vi

return vi.resize(scale)

def fit_tiles(vi, dims=DIMS):

resized dimensions

re_width = vi.width
re_height = vi.height

dims = int(dims)

# forcing perfectly squared tiles on both dimensions
tiles_across_x = re_width // dims
tiles_across_y = re_height // dims

# calculate excess for each dimensions
excess_x = re_width - tiles_across_x * dims
excess_y = re_height - tiles_across_y * dims

return (
    tiles_across_x * dims,
    tiles_across_y * dims,
    excess_x,
    excess_y,
)

def center_crop(vi, coords): fitted_width, fitted_height, excess_x, excess_y = coords

vi = vi.crop(excess_x // 2, excess_y // 2, fitted_width, fitted_height,)

return vi

def deepzoom_save(vi, base_name, dir_name=OUT_DIR, dims=DIMS, file_format=FMT):

as of pyvips 2.1.12: dir_name is required

# and not optional as the docs say
# https://libvips.github.io/pyvips/vimage.html#pyvips.Image.dzsave

vips.Image.dzsave(
    vi,
    dir_name,
    basename=base_name,
    suffix=file_format,
    tile_size=dims,
    overlap=0,
    depth="one",
    properties=False,
)

# delete .dzi file
pathlib.Path(f"{dir_name}/{base_name}.dzi").unlink()

def create_tiles( slide_path, mpp=MPP, tile_dims=DIMS, file_format=FMT, output_dir=OUT_DIR ): t = Timer()

vi = utils.read_image(slide_path)

vi = rescale_slide(vi, mpp=mpp)
coords = fit_tiles(vi, dims=tile_dims)

vi = center_crop(vi, coords)

coords_str = f"{coords[2] // 2}-{coords[3] // 2}"
base_name = f"{slide_path.name}_mpp-{mpp}_crop{coords_str}"

output_path = pathlib.Path(f"{output_dir}/{base_name}_files")
if output_path.is_dir():
    print("- tiles already exist, will skip this slide")
    return output_path

deepzoom_save(
    vi, base_name, dir_name=output_dir, dims=tile_dims, file_format=file_format
)

print(f"- time elapsed: {t.elapsed()}")

return output_path

def save_tiles_csv(output_path, file_format=FMT, relativepaths=True): tiles = [ for _ in output_path.glob(f"0/*{file_format}")] print(f"- directory {output_path}/ has {len(tiles)} {file_format} tiles")

if relative_paths:
    # "output/tiles/CMU-1.svs_mpp-0.5_crop106-72_files/0/31_27.png"
    # .parents[0]: 'output/tiles/CMU-1.svs_mpp-0.5_crop106-72_files/0'
    # .parents[2]: 'output/tiles'
    tiles = [t_path.relative_to(t_path.parents[2]) for t_path in tiles]

tiles_df = pd.DataFrame({"tile_path": tiles})
tiles_csv = f"{output_path.parents[0]}/{output_path.name}-tile_path.csv"
tiles_df.to_csv(tiles_csv, index=False)
print(f"- tile_path data saved to {tiles_csv}")

TODO: extract this function (it is repeated from filters.py)

def consolidate_tiles_csvs(output_dir, consolidated_csv, glob_pattern): output_dir_path = pathlib.Path(output_dir) dfs = [pd.readcsv() for _ in output_dir_path.glob(glob_pattern)] consolidated_df = pd.concat(dfs, ignore_index=True) consolidated_df.to_csv(consolidated_csv, index=False) print(f"tiles info consolidated and saved to {consolidated_csv}")

def run_tiling( slide_csv, consolidated_csv, output_dir=OUT_DIR, mpp=MPP, tile_dims=DIMS, file_format=FMT, ): slides_df = pd.read_csv(slide_csv) slides_df = slides_df.convert_dtypes()

num_slides = len(slides_df)

pathlib.Path(output_dir).mkdir(parents=True, exist_ok=True)

for idx, row in slides_df.iterrows():
    slide_path = pathlib.Path(row["slide_path"])
    print(f"[{idx + 1}/{num_slides}] {slide_path}")

    tiles_path = create_tiles(
        slide_path,
        mpp=mpp,
        tile_dims=tile_dims,
        file_format=file_format,
        output_dir=output_dir,
    )

    save_tiles_csv(tiles_path, file_format=file_format)

consolidate_tiles_csvs(output_dir, consolidated_csv, "*-tile_path.csv")`

Taken from https://github.com/lucasrla/wsi-preprocessing

[manuel-munoz-aguirre]

Thanks for the suggestion! I'll add this to the feature roadmap. Shouldn't be too hard to implement

[kaczmarj]

this feature would be very useful to me, as I am dealing with images that have different base resolutions. let me know if i can help in any way!

[kaczmarj]

@manuel-munoz-aguirre - could you suggest how to implement this feature? i can make a PR but i'm not sure how to best do this.

given an openslide image, would we first try to figure out the best level to use (i.e., get the highest resolution that is close to our target mpp)? once we have that level, we would extract a patch and then resize it to the desired mpp?