cellseg-gsontools 0.1.1

Toolbelt for merging and extracting features from geojson masks.

Cellseg_gsontools

Localized quantification of cell and tissue segmentation maps.

Introduction

Cellseg_gsontools is a Python toolset designed to analyze and summarize large cell and tissue segmentation maps created from Whole Slide Images (WSI). It provides a range of metrics and algorithms out of the box, while also allowing users to define their own functions to meet specific needs. The library is built on top of geopandas and heavily utilizes the GeoDataFrame data structure or gdf for short. In other words, the library is built to process geospatial data with GeoJSON-interface. The library is synergetic with the cellseg_models.pytorch segmentation library which enables you to segment your WSI into GeoJSON format.

NOTE The toolset is still in alpha-phase and under constant development.

Installation

pip install cellseg-gsontools

To add some extra capabilities, like, support for arrow files or abdscan clustering use:

pip install cellseg-gsontools[all]

For latest pre-release, use

pip install cellseg-gsontools==0.1.0a2

Usage

The idea of cellseg_gsontools is to provide an easy-to-use API to extract features from GeoJSON-formatted cell/nuclei/tissue segmentation maps. This can be done via different spatial-analysis methods including:

• Methods for computing morphological metrics.
• Methods for extracting neighborhoods metrics.
• Methods for computing diversity metrics.
• Subsetting cells with tissue areas for more localized feature extraction
• Spatial point clustering methods.
• Utilities for pretty visualization of the spatial data.

Specifically:

• Functional API helps to quickly compute object-level metrics in a GeoDataFrame.
• Spatial Context classes help handling and combining cell and tissue segmentations for more localized and spatially contextualized feature extraction. These classes include algorithms to subset different spatial contexts with methods like spatial joins, graph networks, and clustering.
• Summary classes can be used to reduce context objects into summarised tabular form, if you for some reason happen to be too lazy for geopandas-based data-wrangling. These classes include InstanceSummary, DistanceSummary, SemanticSummary, SpatialWeightSummary

NOTE: The input GeoDataFrames always need to contain a column called class_name or otherwise nothing will work. The column should contain the class or category of the geo-object, e.g. the cell type, or tissue type. This restriction might loosen in the future.

Code examples

Function API

Good to know: These functions can be parallelized through pandarallel by passing in the argument parallel=True.

Geometry

Geometrical (or sometimes morphological) features can be computed for individual nuclear-, cell-, or any type of polygon-objects over the whole GeoDataFrame (gdf). They can be also computed using the summary-object showcased later. These functions take in a gdf and a list of geometrical features and return the same gdf with the computed features for each object.

from cellseg_gsontools.utils import read_gdf
from cellseg_gsontools.geometry import shape_metric

path = "/path/to/cells.json"
gdf = read_gdf(path)
gdf = set_uid(gdf, id_col="uid") # set a running index id column 'uid'

shape_metric(
    gdf,
    metrics = [
        "area",
        "major_axis_len",
        "circularity",
        "eccentricity",
        "squareness"
    ],
    parallel=True
)

uid	geometry	class_name	area	major_axis_len	circularity	eccentricity	squareness
1	Polygon((x, y)...)	inflammatory	161.314012	15.390584	0.942791	0.834633	1.198831
2	Polygon((x, y)...)	connective	401.877306	26.137359	0.948243	0.783638	1.205882
3	Polygon((x, y)...)	connective	406.584839	29.674783	0.877915	0.594909	1.117796
4	Polygon((x, y)...)	connective	281.779998	24.017928	0.885816	0.617262	1.127856
5	Polygon((x, y)...)	connective	257.550056	17.988285	0.891481	0.999339	1.131054
...		...	...	...	...	...	...

Spatial Neighborhood and Neighborhood metrics

Spatial neighborhoods are sets of nodes in large graph networks. The nodes are cell or object-centroids and the links between a node to neighboring nodes define the immediate neighborhood of any object. To extract spatial neighborhoods, a graph needs to be fitted to a gdf. The fit_graph-method can be used to fit a graph network also known as spatial weights object in geospatial analysis terms. Allowed spatial weights fitting methods are ["delaunay", "knn", "distband", "relative_nhood"].

Now if you want to extract features from the immediate neighborhood of objects, you can use the local_character-function. It takes in a gdf and a spatial weights object. The function reduces the neighborhood values to either mean, median or sum. NOTE you can also input a list of columns to the function to compute reductions over many columns.

from cellseg_gsontools.utils import read_gdf, set_uid
from cellseg_gsontools.graphs import fit_graph
from cellseg_gsontools.character import local_character
from cellseg_gsontools.geometry import shape_metric

path = "/path/to/cells.json"
gdf = read_gdf(path)
gdf = set_uid(gdf, id_col="uid") # set a running index id column 'uid'

# compute the eccentricity of the cells
gdf = shape_metric(gdf, metrics = ["eccentricity"], parallel=True)

# fit a spatial weights object
w = fit_graph(gdf, type="delaunay", thresh=150, id_col="uid")

# compute the mean eccentricity of each cell's neighborhood
local_character(
    gdf,
    spatial_weights=w,
    val_col="eccentricity",
    reductions=["mean"], # mean, median, sum,
    weight_by_area=True, # weight the values by the object areas
    parallel=True
)

uid	geometry	class_name	eccentricity	eccentricity_nhood_mean
1	Polygon((x, y)...)	inflammatory	0.556404	0.090223
2	Polygon((x, y)...)	connective	0.474120	0.121348
3	Polygon((x, y)...)	connective	0.712500	0.128495
4	Polygon((x, y)...)	connective	0.325301	0.671348
5	Polygon((x, y)...)	connective	0.285714	0.000000
...		...	...	...

Neighborhood Diversity

Neighborhood-diversity metrics calculate how diverse (or sometimes heterogenous) the immediate neighborhood of an object is. The diversity is computed with respect to a feature (e.g. nuclei area, eccentricity, or class name) of the neighboring objects. Note that the features can also be categorical, like the cell-type class etc. The neighborhood diversities can be computed with the local_diversity-function. The available diversity methods to compute are ["simpson_index", "shannon_index", "gini_index", "theil_index"]. NOTE: you can also input a list of columns to the function to compute diversity metrics over many columns.

from cellseg_gsontools.utils import read_gdf, set_uid
from cellseg_gsontools.graphs import fit_graph
from cellseg_gsontools.geometry import shape_metric
from cellseg_gsontools.diversity import local_diversity

path = "/path/to/cells.json"
gdf = read_gdf(path)
gdf = set_uid(gdf, id_col="cell_id") # set a running index id column 'cell_id'

# compute the eccentricity of the cells
gdf = shape_metric(gdf, metrics = ["area"], parallel=True)

# fit a spatial weights object
w = fit_graph(gdf, type="delaunay", thresh=150, id_col="cell_id")

# compute the heterogeneity of the neighborhood areas
local_diversity(
    gdf,
    spatial_weights=w,
    val_col="area",
    metrics=["simpson_index"],
)

uid	geometry	class_name	area	area_simpson_index
1	Polygon((x, y)...)	inflammatory	161.314012	0.000000
2	Polygon((x, y)...)	connective	401.877306	0.000000
3	Polygon((x, y)...)	connective	406.584839	0.444444
4	Polygon((x, y)...)	connective	281.779998	0.500000
5	Polygon((x, y)...)	connective	257.550056	0.000000
...		...	...	...

Spatial Contexts

Spatial Context classes combine cell segmentation maps with tissue area segmentation maps which helps to extract cells under different spatial contexts. The specific Spatial Context classes are the InterfaceContext WithinContext, and PointClusterContext. All the context-classes include a .fit()-method that builds the context. The .plot()-method can be used to plot figures where the different spatial context areas are highlighted.

WithinContext

Extracts cells from the cell_gdf within areas in area_gdf. Call context2gdf-method to retrieve the cells in a gdf. The different context that can be accessed with the WithinContext-class are ["roi_area", "roi_cells", "roi_network"].

from cellseg_gsontools.spatial_context import WithinContext

within_context = WithinContext(
    area_gdf = area_gdf,
    cell_gdf = cell_gdf,
    label = "area_cin", # Extract the cells that are within tissue areas of this type
    min_area_size = 100000.0 # discard areas smaller than this
)

within_context.fit()
within_context.context2gdf("roi_cells")

InterfaceContext

Returns border region between two types of area given. The different context that can be accessed with the InterfaceContext-class are: ["roi_area", "roi_cells", "roi_network", "interface_cells", "interface_area", "interface_network", "border_network", "full_network"].

from cellseg_gsontools.spatial_context import InterfaceContext

interface_context = InterfaceContext(
    area_gdf = area_gdf,
    cell_gdf = cell_gdf,
    label1 = "area_cin",
    label2 = "areastroma",
    min_area_size = 100000.0 # discard areas smaller than this
)
interface_context.fit()
interface_context.plot(key = "interface_area")

Here we pick the border area between the neoplastic lesion and the stroma to study for example the immune cells on the border.

PointClusterContext

Uses a given clustering algorithm to cluster cells of the given type. This can help to extract spatial clusters of cells.

from cellseg_gsontools.spatial_context import PointClusterContext

cluster_context = PointClusterContext(
    cell_gdf = cell_gdf,
    label = "inflammatory", # cells of this type will be clustered
    cluster_method = "dbscan",
    min_area_size = 5000.0, # dbscan param
    min_samples = 70 # dbscan param
)

cluster_context.fit()
cluster_context.plot_weights("roi_network")

Here we clustered the immune cells on the slide and fitted a network on the cells that were within the alpha shape of the cluster.

Summary

Summarize cells, areas, contexts, and intermediates of a slide into a tabular format for further analysis. Summarise-method must be called before using summary. The summaries can be grouped by any metadata or annotation column in the gdf. You can also use filter_pattern argument to choose the statistics (mean, count etc.) or groups used in summary output.

InstanceSummary

Easy way to calculate nuclei and area metrics for different classes of cells over the neoplastic areas of the tissue.

neoplastic_areas = within_context.context2gdf("roi_cells")
spatial_weights =  within_context.context2weights("interface_cells", threshold=75.0)

lesion_summary = InstanceSummary(
    neoplastic_areas,
    metrics = ["area"],
    groups = ["class_name"],
    spatial_weights = spatial_weights,
    prefix = "lesion-cells-"
)
lesion_summary.summarize()

	sample_cells
lesion-cells-inflammatory-count	118.00
lesion-cells-neoplastic-count	4536.00
lesion-cells-total-count	4787.00
lesion-cells-inflammatory-area-mean	241.17
lesion-cells-neoplastic-area-mean	532.79

SemanticSummary

Summarizes tissues areas. Here we summarize the areas of immune clusters in the while tissue.

immune_cluster_areas = cluster_context.context2gdf("roi_area")
immune_areas = SemanticSummary(
    immune_cluster_areas,
    metrics = ["area"],
    prefix = "immune-clusters-"
)
immune_areas.summarize()

	sample_cells
immune-clusters-immune-clusters-total-count	42.00
immune-clusters-area-mean	2558514.25

SpatialWeightSummary

Summarizes cell networks by counting edges between neighboring cells. Here we compute the cell-cell connections over the tumor-stroma interface. The cell-cell connections are defined by the spatial weights graph.

interface_summary = SpatialWeightSummary(
    iface_context.merge_weights("border_network"),
    iface_context.cell_gdf,
    classes= ["inflammatory", "neoplastic"],
    prefix = "interface-"
)
interface_summary.summarize()

	sample_cells
interface-inflammatory-inflammatory	19
interface-inflammatory-neoplastic	153
interface-neoplastic-neoplastic	363

DistanceSummary

Summarizes distances between different areas. For example how many immune clusters are close to a neoplastic lesion.

immune_proximities = DistanceSummary(
    immune_cluster_areas,
    neoplastic_areas,
    groups = None,
    prefix = "icc-close2lesion-",
)
immune_proximities.summarize()

	sample_cells
icc-close2lesion-0-count	34
icc-close2lesion-1-count	8