Getting Started🔗
The large_image library can be used to read and access different file formats. There are several common usage patterns.
To read more about accepted file formats, visit the Image Formats page.
These examples use sample.tiff as an example – any readable image can be used in this case. Visit demo.kitware.com to download a sample image.
Installation🔗
In addition to installing the base large-image package, you’ll need at least one tile source which corresponds to your target file format(s) (a large-image-source-xxx package). You can install everything from the main project with one of these commands:
Pip🔗
Install common tile sources on linux, OSX, or Windows:
pip install large-image[common]
Install all tile sources on linux:
pip install large-image[all] --find-links https://girder.github.io/large_image_wheels
When using large-image with an instance of Girder, install all tile sources and all Girder plugins on linux:
pip install large-image[all] girder-large-image-annotation[tasks] --find-links https://girder.github.io/large_image_wheels
Conda🔗
Conda makes dependency management a bit easier if not on Linux. The base module, converter module, and two of the source modules are available on conda-forge. You can install the following:
conda install -c conda-forge large-image
conda install -c conda-forge large-image-source-gdal
conda install -c conda-forge large-image-source-tiff
conda install -c conda-forge large-image-converter
Reading Image Metadata🔗
All images have metadata that include the base image size, the base tile size, the number of conceptual levels, and information about the size of a pixel in the image if it is known.
import large_image
source = large_image.open('sample.tiff')
print(source.getMetadata())
This might print a result like:
{
'levels': 9,
'sizeX': 58368,
'sizeY': 12288,
'tileWidth': 256,
'tileHeight': 256,
'magnification': 40.0,
'mm_x': 0.00025,
'mm_y': 0.00025
}
levels doesn’t actually tell which resolutions are present in the file. It is the number of levels that can be requested from the getTile method. The levels can also be computed via ceil(log(max(sizeX / tileWidth, sizeY / tileHeight)) / log(2)) + 1.
The mm_x and mm_y values are the size of a pixel in millimeters. These can be None if the value is unknown. The magnification is that reported by the file itself, and may be None. The magnification can be approximated by 0.01 / mm_x.
Getting a Region of an Image🔗
You can get a portion of an image at different resolutions and in different formats. Internally, the large_image library reads the minimum amount of the file necessary to return the requested data, caching partial results in many instances so that a subsequent query may be faster.
import large_image
source = large_image.open('sample.tiff')
image, mime_type = source.getRegion(
region=dict(left=1000, top=500, right=11000, bottom=1500),
output=dict(maxWidth=1000),
encoding='PNG')
# image is a PNG that is 1000 x 100. Specifically, it will be a bytes
# object that represent a PNG encoded image.
You could also get this as a numpy array:
import large_image
source = large_image.open('sample.tiff')
nparray, mime_type = source.getRegion(
region=dict(left=1000, top=500, right=11000, bottom=1500),
output=dict(maxWidth=1000),
format=large_image.constants.TILE_FORMAT_NUMPY)
# Our source image happens to be RGB, so nparray is a numpy array of shape
# (100, 1000, 3)
You can specify the size in physical coordinates:
import large_image
source = large_image.open('sample.tiff')
nparray, mime_type = source.getRegion(
region=dict(left=0.25, top=0.125, right=2.75, bottom=0.375, units='mm'),
scale=dict(mm_x=0.0025),
format=large_image.constants.TILE_FORMAT_NUMPY)
# Since our source image had mm_x = 0.00025 for its scale, this has the
# same result as the previous example.
Tile Serving🔗
One of the uses of large_image is to get tiles that can be used in image or map viewers. Most of these viewers expect tiles that are a fixed size and known resolution. The getTile method returns tiles as stored in the original image and the original tile size. If there are missing levels, these are synthesized – this is only done for missing powers-of-two levels or missing tiles. For instance,
import large_image
source = large_image.open('sample.tiff')
# getTile takes x, y, z, where x and y are the tile location within the
# level and z is level where 0 is the lowest resolution.
tile0 = source.getTile(0, 0, 0)
# tile0 is the lowest resolution tile that shows the whole image. It will
# be a JPEG or PNG or some other image format depending on the source
tile002 = source.getTile(0, 0, 2)
# tile002 will be a tile representing no more than 1/4 the width of the
# image in the upper-left corner. Since the z (third parameter) is 2, the
# level will have up to 2**2 x 2**2 (4 x 4) tiles. An image doesn't
# necessarily have all tiles in that range, as the image may not be square.
Some methods such as getRegion and getThumbnail allow you to specify format on the fly. But note that since tiles need to be cached in a consistent format, getTile always returns the same format depending on what encoding was specified when it was opened:
import large_image
source = large_image.open('sample.tiff', encoding='PNG')
tile0 = source.getTile(0, 0, 0)
# tile is now guaranteed to be a PNG
Tiles are always tileWidth by tileHeight in pixels. At the maximum level (z = levels - 1), the number of tiles in that level will range in x from 0 to strictly less than sizeX / tileWidth, and y from 0 to strictly less than sizeY / tileHeight. For each lower level, the is a power of two less tiles. For instance, when z = levels - 2, x ranges from 0 to less than sizeX / tileWidth / 2; at z = levels - 3, x is less than sizeX / tileWidth / 4.
Iterating Across an Image🔗
Since most images are too large to conveniently fit in memory, it is useful to iterate through the image.
The tileIterator function can take the same parameters as getRegion to pick an output size and scale, but can also specify a tile size and overlap.
You can also get a specific tile with those parameters. This tiling doesn’t have to have any correspondence to the tiling of the original file.
The data for each tile is loaded lazily, only once tile['tile'] or tile['format'] is accessed.
import large_image
source = large_image.open('sample.tiff')
for tile in source.tileIterator(
tile_size=dict(width=512, height=512),
format=large_image.constants.TILE_FORMAT_NUMPY
):
# tile is a dictionary of information about the specific tile
# tile['tile'] contains the actual numpy or image data
print(tile['x'], tile['y'], tile['tile'].shape)
# This will print something like:
# 0 0 (512, 512, 3)
# 512 0 (512, 512, 3)
# 1024 0 (512, 512, 3)
# ...
# 56832 11776 (512, 512, 3)
# 57344 11776 (512, 512, 3)
# 57856 11776 (512, 512, 3)
You can overlap tiles. For instance, if you are running an algorithm where there are edge effects, you probably want an overlap that is big enough that you can trim off or ignore those effects:
import large_image
source = large_image.open('sample.tiff')
for tile in source.tileIterator(
tile_size=dict(width=2048, height=2048),
tile_overlap=dict(x=128, y=128, edges=False),
format=large_image.constants.TILE_FORMAT_NUMPY
):
print(tile['x'], tile['y'], tile['tile'].shape)
# This will print something like:
# 0 0 (2048, 2048, 3)
# 1920 0 (2048, 2048, 3)
# 3840 0 (2048, 2048, 3)
# ...
# 53760 11520 (768, 2048, 3)
# 55680 11520 (768, 2048, 3)
# 57600 11520 (768, 768, 3)
Getting a Thumbnail🔗
You can get a thumbnail of an image in different formats or resolutions. The default is typically JPEG and no larger than 256 x 256. Getting a thumbnail is essentially the same as doing getRegion, except that it always uses the entire image and has a maximum width and/or height.
import large_image
source = large_image.open('sample.tiff')
image, mime_type = source.getThumbnail()
open('thumb.jpg', 'wb').write(image)
You can get the thumbnail in other image formats and sizes:
import large_image
source = large_image.open('sample.tiff')
image, mime_type = source.getThumbnail(width=640, height=480, encoding='PNG')
open('thumb.png', 'wb').write(image)
Associated Images🔗
Many digital pathology images (also called whole slide images or WSI) contain secondary images that have additional information. This commonly includes label and macro images. A label image is a separate image of just the label of a slide. A macro image is a small image of the entire slide either including or excluding the label. There can be other associated images, too.
import large_image
source = large_image.open('sample.tiff')
print(source.getAssociatedImagesList())
# This prints something like:
# ['label', 'macro']
image, mime_type = source.getAssociatedImage('macro')
# image is a binary image, such as a JPEG
image, mime_type = source.getAssociatedImage('macro', encoding='PNG')
# image is now a PNG
image, mime_type = source.getAssociatedImage('macro', format=large_image.constants.TILE_FORMAT_NUMPY)
# image is now a numpy array
You can get associated images in different encodings and formats. The entire image is always returned.
Projections🔗
large_image handles geospatial images. These can be handled as any other image in pixel-space by just opening them normally. Alternately, these can be opened with a new projection and then referenced using that projection.
import large_image
# Open in Web Mercator projection
source = large_image.open('sample.geo.tiff', projection='EPSG:3857')
print(source.getMetadata()['bounds'])
# This will have the corners in Web Mercator meters, the projection, and
# the minimum and maximum ranges.
# We could also have done
print(source.getBounds())
# The 0, 0, 0 tile is now the whole world excepting the poles
tile0 = source.getTile(0, 0, 0)
Images with Multiple Frames🔗
Some images have multiple “frames”. Conceptually, these are images that could have multiple channels as separate images, such as those from fluorescence microscopy, multiple “z” values from serial sectioning of thick tissue or adjustment of focal plane in a microscope, multiple time (“t”) values, or multiple regions of interest (frequently referred as “xy”, “p”, or “v” values).
Any of the frames of such an image are accessed by adding a frame=<integer> parameter to the getTile, getRegion, tileIterator, or other methods.
import large_image
source = large_image.open('sample.ome.tiff')
print(source.getMetadata())
# This will print something like
# {
# 'magnification': 8.130081300813009,
# 'mm_x': 0.00123,
# 'mm_y': 0.00123,
# 'sizeX': 2106,
# 'sizeY': 2016,
# 'tileHeight': 1024,
# 'tileWidth': 1024,
# 'IndexRange': {'IndexC': 3},
# 'IndexStride': {'IndexC': 1},
# 'frames': [
# {'Frame': 0, 'Index': 0, 'IndexC': 0, 'IndexT': 0, 'IndexZ': 0},
# {'Frame': 1, 'Index': 0, 'IndexC': 1, 'IndexT': 0, 'IndexZ': 0},
# {'Frame': 2, 'Index': 0, 'IndexC': 2, 'IndexT': 0, 'IndexZ': 0}
# ]
# }
nparray, mime_type = source.getRegion(
frame=1,
format=large_image.constants.TILE_FORMAT_NUMPY)
# nparray will contain data from the middle channel image
Channels, Bands, Samples, and Axes🔗
Various large image formats refer to channels, bands, and samples. This isn’t consistent across different libraries. In an attempt to harmonize the geospatial and medical image terminology, large_image uses bands or samples to refer to image plane components, such as red, green, blue, and alpha. For geospatial data this can often have additional bands, such as near infrared or panchromatic. channels are stored as separate frames and can be interpreted as different imaging modalities. For example, a fluorescence microscopy image might have DAPI, CY5, and A594 channels. A common color photograph file has 3 bands (also called samples) and 1 channel.
At times, image axes are used to indicate the order of data, especially when interpreted as an n-dimensional array. The x and y axes are the horizontal and vertical dimensions of the image. The s axis is the bands or samples, such as red, green, and blue. The c axis is the channels with special support for channel names. This corresponds to distinct frames.
The z and t are common enough that they are sometimes considered as primary axes. z corresponds to the direction orthogonal to x and y and is usually associated with altitude or microscope stage height. t is time.
Other axes are supported provided their names are case-insensitively unique.
Styles - Changing colors, scales, and other properties🔗
By default, reading from an image gets the values stored in the image file. If you get a JPEG or PNG as the output, the values will be 8-bit per channel. If you get values as a numpy array, they will have their original resolution. Depending on the source image, this could be 16-bit per channel, floats, or other data types.
Especially when working with high bit-depth images, it can be useful to modify the output. For example, you can adjust the color range:
import large_image
source = large_image.open('sample.tiff', style={'min': 'min', 'max': 'max'})
# now, any calls to getRegion, getTile, tileIterator, etc. will adjust the
# intensity so that the lowest value is mapped to black and the brightest
# value is mapped to white.
image, mime_type = source.getRegion(
region=dict(left=1000, top=500, right=11000, bottom=1500),
output=dict(maxWidth=1000))
# image will use the full dynamic range
You can also composite a multi-frame image into a false-color output:
import large_image
source = large_image.open('sample.tiff', style={'bands': [
{'frame': 0, 'min': 'min', 'max': 'max', 'palette': '#f00'},
{'frame': 3, 'min': 'min', 'max': 'max', 'palette': '#0f0'},
{'frame': 4, 'min': 'min', 'max': 'max', 'palette': '#00f'},
]})
# Composite frames 0, 3, and 4 to red, green, and blue channels.
image, mime_type = source.getRegion(
region=dict(left=1000, top=500, right=11000, bottom=1500),
output=dict(maxWidth=1000))
# image is false-color and full dynamic range of specific frames
Writing an Image🔗
If you wish to visualize numpy data, large_image can write a tiled tiff. This requires a tile source that supports writing to be installed. As of this writing, the large-image-source-zarr and large-image-source-vips sources supports this. If both are installed, the large-image-source-zarr is the default.
import large_image
source = large_image.new()
for nparray, x, y in fancy_algorithm():
# We could optionally add a mask to limit the output
source.addTile(nparray, x, y)
source.write('/tmp/sample.tiff', lossy=False)
The large-image-source-zarr can be used to store multiple frame data with arbitrary axes.
import large_image
source = large_image.new()
for nparray, x, y, time, param1 in fancy_algorithm():
source.addTile(nparray, x, y, time=time, p1=param1)
# The writer supports a variety of formats
source.write('/tmp/sample.zarr.zip', lossy=False)
You may also choose to read tiles from one source and write modified tiles to a new source:
import large_image
original_source = large_image.open('path/to/original/image.tiff')
new_source = large_image.new()
for frame in original_source.getMetadata().get('frames', []):
for tile in original_source.tileIterator(frame=frame['Frame'], format='numpy'):
t, x, y = tile['tile'], tile['x'], tile['y']
kwargs = {
'z': frame['IndexZ'],
'c': frame['IndexC'],
}
modified_tile = modify_tile(t)
new_source.addTile(modified_tile, x=x, y=y, **kwargs)
new_source.write('path/to/new/image.tiff', lossy=False)
In some cases, it may be beneficial to write to a single image from multiple processes or threads:
import large_image
import multiprocessing
# Important: Must be a pickleable function
def add_tile_to_source(tilesource, nparray, position):
tilesource.addTile(
nparray,
**position
)
source = large_image.new()
# Important: Maximum size must be allocated before any multiprocess concurrency
add_tile_to_source(source, np.zeros(1, 1, 3), dict(x=max_x, y=max_y, z=max_z))
# Also works with multiprocessing.ThreadPool, which does not need maximum size allocated first
with multiprocessing.Pool(max_workers=5) as pool:
pool.starmap(
add_tile_to_source,
[(source, t, t_pos) for t, t_pos in tileset]
)
source.write('/tmp/sample.zarr.zip', lossy=False)