pythologist
Read and analyze cell image data.
Intro
Pythologist 1) reads exports from InForm software or other sources into a common storage format, and 2) extracts basic analysis features from cell image data. This software is generally intended to be run from a jupyter notebook and provides hooks into the image data so that the user can have the flexability to execute analyses they design or find in the primary literature.
List of large-scale image analysis publications
Pythologist is based on IrisSpatialFeatures (C.D. Carey, ,D. Gusenleitner, M. Lipshitz, et al. Blood. 2017) https://doi.org/10.1182/blood-2017-03-770719, and is implemented in the python programming language.
Features Pythologist add are:
- An common CellProjectGeneric storage class, and classical inheritance conventions to organize the importation of different data types.
- A mutable CellDataFrame class that can be used for slicing, and combining projects.
- The ability to add binary features to cells based on cell-cell contacts or cell proximity.
- Customizable images based on the cell segmentation or heatmaps spaninng the cartesian coordinates.
- Specify cell populations through a SubsetLogic syntax for quick selection of mutually exclusive phenotypes or binary features
- A set of Quality Check functions to identify potential issues in imported data.
Module documentation
-
pythologist
CellDataFrame class to modify and execute analysses [Read the Docs] [source] -
pythologist-reader
CellProject Storage Object [Read the Docs] [source] -
pythologist-test-images
Example data [source] -
pythologist-image-utilities
Helper functions to work with images [Read the Docs] [source]
Quickstart
To start a jupyter lab notebook with the required software as your user in your current drectory you can use the following command
docker run --rm -p 8888:8888 --user $(id -u):$(id -g) -v $(pwd):/work vacation/pythologist:latest
This will start jupyter lab on port 8888 as your user and group.
Any of the test data examples should work fine in this environment.
Installation
Docker jupyter labs quickstart
To start a jupyter notebook in the current working directory on port 8885 you can use the following docker command.
First build a docker image that will use your own user/group name and id.
$ docker build -t pythologist:latest --build-arg user=$USERNAME \
--build-arg group=$GROUPNAME \
--build-arg user_id=$USERID \
--build-arg group_id=$GROUPID .
Now start the docker image.
$ docker run --rm -p 8885:8888 -v $(pwd):/home/$USERNAME/work pythologist:latest
where $USERNAME
, $GROUPNAME
, $USERID
, $GROUPID
correspond to your user/group name/id.
Install by pip
$ pip install pythologist
Common tasks
Reading in a project composed of InForm exports
Basic
The assumption here is that the exports are grouped so that sample folders contain one or more image exports, and that sample name can be inferred from the last folder name.
from pythologist_test_images import TestImages
from pythologist_reader.formats.inform.sets import CellProjectInForm
import matplotlib.pyplot as plt
# Get the path of the test dataset
path = TestImages().raw('IrisSpatialFeatures')
# Create the storage opbject where the project will be saved
cpi = CellProjectInForm('pythologist.h5',mode='w')
# Read the project data
cpi.read_path(path,require=False,verbose=True,microns_per_pixel=0.496,sample_name_index=-1)
# Display one of the cell map images
for f in cpi.frame_iter():
break
print(f.frame_name)
plt.imshow(f.cell_map_image(),origin='upper')
plt.show()
MEL2_7
Custom region annotations from tumor and invasive margin image drawings
Another format supported for a project import is one with a custom tumor and invasive margin definition. Similar to above, the project is organized into sample folders, and each image within each sample folder has a tif file defining the tumor and invasive margin. These come in the form of a <image name prefix>_Tumor.tif
and <image name prefix>_Invasive_Margin.tif
for each image. The _Tumor.tif
is an area filled in where the tumor is, and transparent elsewhere. The _Invasive_Margin.tif
is a drawn line of a known width. steps
is used to grow the margin out that many pixels in each direction to establish an invasive margin region. Here we also rename some markers during read-in to clean up the syntax of thresholding on binary features.
from pythologist_test_images import TestImages
from pythologist_reader.formats.inform.custom import CellProjectInFormLineArea
# Get the path of the test dataset
path = TestImages().raw('IrisSpatialFeatures')
# Specify where the data read-in will be stored as an h5 object
cpi = CellProjectInFormLineArea('test.h5',mode='w')
# Read in the data (gets stored on the fly into the h5 object)
cpi.read_path(path,
sample_name_index=-1,
verbose=True,
steps=76,
project_name='IrisSpatialFeatures',
microns_per_pixel=0.496)
for f in cpi.frame_iter():
break
print(f.frame_name)
print('hand drawn margin')
plt.imshow(f.get_image(f.get_data('custom_images').\
set_index('custom_label').loc['Drawn','image_id']),origin='upper')
plt.show()
print('hand drawn tumor area')
plt.imshow(f.get_image(f.get_data('custom_images').\
set_index('custom_label').loc['Area','image_id']),origin='upper')
plt.show()
print('Mutually exclusive Margin, Tumor, and Stroma')
plt.imshow(f.get_image(f.get_data('regions').\
set_index('region_label').loc['Margin','image_id']),origin='upper')
plt.show()
plt.imshow(f.get_image(f.get_data('regions').\
set_index('region_label').loc['Tumor','image_id']),origin='upper')
plt.show()
plt.imshow(f.get_image(f.get_data('regions').\
set_index('region_label').loc['Stroma','image_id']),origin='upper')
plt.show()
MEL2_2
hand drawn margin
hand drawn tumor area
Mutually exclusive Margin, Tumor, and Stroma
Read a project with a custom tumor mask (but no margin line)
Here we will use the mask, but not expand or subtract from it.
from pythologist_test_images import TestImages
from pythologist_reader.formats.inform.custom import CellProjectInFormCustomMask
import matplotlib.pyplot as plt
path = TestImages().raw('IrisSpatialFeatures')
cpi = CellProjectInFormCustomMask('test.h5',mode='w')
cpi.read_path(path,
microns_per_pixel=0.496,
sample_name_index=-1,
verbose=True,
custom_mask_name='Tumor',
other_mask_name='Not-Tumor')
for f in cpi.frame_iter():
rs = f.get_data('regions').set_index('region_label')
for r in rs.index:
print(r)
plt.imshow(f.get_image(rs.loc[r]['image_id']),origin='upper')
plt.show()
break
MEL2_2
Tumor
Not-Tumor
Quality check samples
Check general status of the CellDataFrame
cdf = cpi.cdf
cdf.db = cpi
cdf.qc(verbose=True).print_results()
prints the following QC metrics to stdout
==========
Check microns per pixel attribute
PASS
Microns per pixel is 0.496
==========
Check storage object is set
PASS
h5 object is set
==========
Is there a 1:1 correspondence between sample_name and sample_id?
PASS
Good concordance.
Issue count: 0/2
==========
Is there a 1:1 correspondence between frame_name and frame_id?
PASS
Good concordance.
Issue count: 0/4
==========
Is there a 1:1 correspondence between project_name and project_id?
PASS
Good concordance.
Issue count: 0/1
==========
Is the same frame name present in multiple samples?
PASS
frame_name's are all in their own samples
Issue count: 0/4
==========
Are the same phenotypes listed and following rules for mutual exclusion?
PASS
phenotype_calls and phenotype_label follows expected rules
==========
Are the same phenotypes included on all images?
PASS
Consistent phenotypes
Issue count: 0/4
==========
Are the same scored names included on all images?
PASS
Consistent scored_names
Issue count: 0/4
==========
Are the same regions represented the same with an image and across images?
PASS
Consistent regions
Issue count: 0/5
==========
Are the same regions listed matching a valid region_label
PASS
regions and region_label follows expected rules
==========
Do we have any region sizes so small they should consider being excluded?
WARNING
[
"Very small non-zero regions are included in the data['IrisSpatialFeatures', 'MEL2', 'MEL2_7', {'Margin': 495640, 'Tumor': 947369, 'Stroma': 116}]"
]
Issue count: 1/2
View density plots based on cell phenotype frequencies.
The cell phenotypes set prior to calling cartesian
are the phenotypes available to plot.
from pythologist_test_images import TestImages
from plotnine import *
proj = TestImages().project('IrisSpatialFeatures')
cdf = TestImages().celldataframe('IrisSpatialFeatures')
cdf.db = proj
cart = cdf.cartesian(verbose=True,step_pixels=50,max_distance_pixels=75)
df,cols = cart.rgb_dataframe(red='CD8+',green='SOX10+')
shape = cdf.iloc[0]['frame_shape']
(ggplot(df,aes(x='frame_x',y='frame_y',fill='color_str'))
+ geom_point(shape='h',size=4.5,color='#777777',stroke=0.2)
+ geom_vline(xintercept=-1,color="#555555")
+ geom_vline(xintercept=shape[1],color="#555555")
+ geom_hline(yintercept=-1,color="#555555")
+ geom_hline(yintercept=shape[0],color="#555555")
+ facet_wrap('frame_name')
+ scale_fill_manual(cols,guide=False)
+ theme_bw()
+ theme(figure_size=(8,8))
+ theme(aspect_ratio=shape[0]/shape[1])
+ scale_y_reverse()
)
View histograms of pixel intensity and the scoring of binary markers on each image
from pythologist_test_images import TestImages
from plotnine import *
proj = TestImages().project('IrisSpatialFeatures')
cdf = TestImages().celldataframe('IrisSpatialFeatures')
cdf.db = proj
ch = cdf.db.qc().channel_histograms()
sub = ch.loc[(~ch['threshold_value'].isna())&(ch['channel_label']=='PDL1')]
(ggplot(sub,aes(x='bins',y='counts'))
+ geom_bar(stat='identity')
+ facet_wrap('frame_name')
+ geom_vline(aes(xintercept='threshold_value'),color='red')
+ theme_bw()
+ ggtitle('Thresholding of PDL1\ngiven image pixel intensities')
)
The original component images were not available for IrisSpatialFeatures example, so pixel intensities are simulated and don't necessarily match the those which would have been used to set the original threshold values.
View cell-cell contacts
from pythologist_test_images import TestImages
from pythologist_reader.formats.inform.custom import CellProjectInFormCustomMask
from pythologist import SubsetLogic as SL
cpi = TestImages().project('IrisSpatialFeatures')
cdf = cpi.cdf
cdf.db = cpi
sub = cdf.loc[cdf['frame_name']=='MEL2_7'].dropna()
cont = sub.contacts().threshold('CD8+','CD8+/contact').contacts().threshold('SOX10+','SOX10+/contact')
cont = cont.threshold('CD8+','SOX10+/contact',
positive_label='CD8+ contact',
negative_label='CD8+').\
threshold('SOX10+','CD8+/contact',
positive_label='SOX10+ contact',
negative_label='SOX10+')
schema = [
{'subset_logic':SL(phenotypes=['OTHER']),
'edge_color':(50,50,50,255),
'watershed_steps':0,
'fill_color':(0,0,0,255)
},
{'subset_logic':SL(phenotypes=['SOX10+']),
'edge_color':(166,206,227,255),
'watershed_steps':0,
'fill_color':(0,0,0,0)
},
{'subset_logic':SL(phenotypes=['CD8+']),
'edge_color':(253,191,111,255),
'watershed_steps':0,
'fill_color':(0,0,0,0)
},
{'subset_logic':SL(phenotypes=['CD8+ contact']),
'edge_color':(253,191,111,255),
'watershed_steps':0,
'fill_color':(255,127,0,255)
},
{'subset_logic':SL(phenotypes=['SOX10+ contact']),
'edge_color':(166,206,227,255),
'watershed_steps':0,
'fill_color':(31,120,180,255)
}
]
sio = cont.segmentation_images().build_segmentation_image(schema,background=(0,0,0,255))
sio.write_to_path('test_edges',overwrite=True)
MEL2_7
Image is zoomed-in and cropped to show the contours better.
Merge CellDataFrames that have the same image segmentations but different scored calls
This happens frequently because current InForm exports only permit two features to be scored per export
merged,fail = cdf1.merge_scores(cdf2,on=['sample_name','frame_name','x','y'])
Show names of the binary 'scored_calls'
cdf.scored_names
['PD1', 'PDL1']
Show phenotypes
cdf.phenotypes
['CD8+', 'OTHER', 'SOX10+']
Show regions
cdf.regions
['Margin', 'Stroma', 'Tumor']
Combine two or more phenotypes into one or rename a phenotype
collapsed = cdf.collapse_phenotypes(['CD8+','OTHER'],'non-Tumor')
collapsed.phenotypes
['SOX10+', 'non-Tumor']
Rename a region
Rename TUMOR to Tumor
renamed = cdf.rename_region('TUMOR','Tumor')
Rename scored phenotypes
renamed = cdf.rename_scored_calls({'PDL1 (Opal 520)':'PDL1'})
Threshold a phenotype
Make CYTOK into CYTOK PDL1+ and CYTOK PDL1-
raw_thresh = raw.threshold('CYTOK','PDL1')
Double threshold
CD68_CD163 = raw.threshold('CD68','CD163').\
threshold('CD68 CD163+','PDL1').\
threshold('CD68 CD163-','PDL1')
Get per frame counts
generate counts and fractions of the current phenotypes and export them to a csv
cdf.counts().frame_counts().to_csv('my_frame_counts.csv')
Get per sample counts
generate counts and fractions of the current phenotypes and export them to a csv
cdf.counts().sample_counts().to_csv('my_sample_counts.csv')
Identify cells that are in contact with a phenotype
The follow command creates a new CellDataFrame that has an additional binary feature representative of the contact with 'T cell' phenotype cells.
cdf = cdf.contacts().threshold('T cell')
Identify cells that are within some proximity of a phenotype of interest
The follow command creates a new CellDataFrame that has an additional binary feature representative of being inside or outisde 75 microns of 'T cell' phenotype cells.
cdf = cdf.nearestneighbors().threshold('T cell','T cell/within 75um',distance_um=75)
Create an image of cell-cell contacts between features of interest
Check outputs against IrisSpatialFeatures outputs
To ensure we are generating expected outs we can check against the outputs of IrisSpatialFeatures [github].
- Jupyter Notebook: Test against IrisSpatialFeatures outputs