Spatially aware cell cluster (SPACCL) graphs

Patent metadata

Field	Value
Publication number	US-9430830-B2
Application number	US-201414564298-A
Country	US
Kind code	B2
Filing date	Dec 9, 2014
Priority date	Jan 30, 2014
Publication date	Aug 30, 2016
Grant date	Aug 30, 2016

Abstract

Methods, apparatus, and other embodiments associated with objectively predicting disease aggressiveness using Spatially Aware Cell Cluster (SpACCl) graphs. One example apparatus includes a set of logics that acquires an image of a region of tissue, partitions the image into a stromal compartment and an epithelial compartment, identifies cluster nodes within the compartments, constructs a spatially aware stromal sub-graph and a spatially aware epithelial sub-graph based on the cluster nodes and a probabilistic decaying function of the distance between cluster nodes, extracts local features from the sub-graphs, and predicts the aggressiveness of a disease in the region of tissue based on the sub-graphs and the extracted features. Example methods and apparatus may employ a Support Vector Machine classifier to classify super-pixels within the image as stromal super-pixels or epithelial super-pixels.

First claim

What is claimed is:

1. A non-transitory, computer-readable storage medium storing computer-executable instructions that when executed by the computer control the computer to perform a method for predicting a risk of disease by examining architectural features of stromal and epithelial tissue with a spatially aware cell cluster graph (SpACCl), comprising:
accessing an image of a region of pathological tissue;
identifying a stromal compartment in the image;
identifying an epithelial compartment in the image, where the epithelial compartment is distinguishable from the stromal compartment;
identifying a plurality of cluster nodes in the image, where a cluster node comprises a plurality of nuclei, and where identifying a plurality of cluster nodes comprises:
identifying a stromal cluster node in the stromal compartment, and
identifying an epithelial cluster node in the epithelial compartment;

constructing electronic data associated with a spatially aware stromal sub-graph G S by connecting a first stromal cluster node with a second, different stromal cluster node, where the probability that the first stromal cluster node is connected with the second stromal cluster node is based, at least in part, on a probabilistic decaying function of the relative distance between the first stromal cluster node and the second stromal cluster node;
constructing a spatially aware epithelial sub-graph G E by connecting a first epithelial cluster node with a second, different epithelial cluster node, where the probability that the first epithelial cluster node is connected with the second epithelial cluster node is based, at least in part, on a probabilistic decaying function of the relative distance between the first epithelial cluster node and the second epithelial cluster node;
extracting local graph features from the sub-graphs G S and G E ; and
predicting the risk of disease, based, at least in part, on the local graph features.

2. The non-transitory, computer-readable storage medium of claim 1 , where identifying a stromal compartment in the image and identifying an epithelial compartment in the image comprises:
partitioning the image into a plurality of spatially coherent super-pixels;
identifying nuclei within a super-pixel;
generating a set of measurements by measuring the intensity and texture of the super-pixel and a neighboring super-pixel;
training a classifier on the set of measurements;
using the classifier to classify the super-pixel as either a stromal super-pixel or epithelial super-pixel;
upon determining that the super-pixel is a stromal super-pixel, assigning the stromal super-pixel to the stromal compartment; and
upon determining that the super-pixel is an epithelial super-pixel, assigning the epithelial super-pixel to the epithelial compartment.

3. The non-transitory, computer-readable storage medium of claim 2 , where the classifier is a Support Vector Machine (SVM) classifier, and where the SVM classifier is trained on the set of measurements using hand-labelled super-pixels.

4. The non-transitory, computer-readable storage medium of claim 1 , where identifying a plurality of cluster nodes in the image comprises:
sampling three consecutive points (c w−1 , c w , c w+1 ) on a contour;
computing an angle θ(c w ) between a plurality of vectors, where the plurality of vectors is defined by sampling the three consecutive points on the contour;
determining a degree of concavity, where the degree of concavity is proportional to the angle θ(c w );
designating a point as a concavity point if θ(c w )>θ t , where θ t is an empirically set threshold degree;
calculating a number of concavity points, and
upon determining that the number of concavity points c w ≧1,
classifying the contour as a cluster node.

5. The non-transitory, computer-readable storage medium of claim 1 , where the probabilistic decaying function of the relative distance between the first stromal cluster node and the second stromal cluster node is defined as:
P ( u,v )= d ( u,v ) −α ,
where
u represents the first stromal cluster node,
v represents the second stromal cluster node, and
0≦α.

6. The non-transitory, computer-readable storage medium of claim 1 , where the probabilistic decaying function of the relative distance between the first epithelial cluster node and the second epithelial cluster node is defined as:
P ( u,v )= d ( u,v ) −α ,
where
u represents the first epithelial cluster node,
v represents the second epithelial cluster node, and
0≦α.

7. The non-transitory, computer-readable storage medium of claim 6 , where a set of edges E i in the sub-graph G S or the sub-graph G E is defined as E i ={(u, v): r<d(u, v) −α , ∀u, vεV i }, where r is a real number between 0 and 1, and where α controls the density of the sub-graph.

8. The non-transitory, computer-readable storage medium of claim 7 , where extracting local graph features from the sub-graph G S and the sub-graph G E comprises extracting a clustering coefficient C, a clustering coefficient D, a giant connected component, an average eccentricity, a percent of isolated points, a number of central points, or a skewness of edge lengths.

9. The non-transitory, computer readable storage medium of claim 8 , where the clustering coefficient C describes a ratio of a total number of edges among neighbors of a cluster node to a total maximum possible number of edges among neighbors of the cluster node, per cluster node, where the clustering coefficient C is defined as:

C
~

=

∑

u
=
1


V


⁢

C
u


V


,

where
⁢

⁢

C
u

=



E
u



(

k
u

2

)

=

2
⁢



E
u

Assignees

• Univ Case Western Reserve

Inventors

• Madabhushi Anant
• Ali Sahirzeeshan

Classifications

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/758

  Involving statistics of pixels or of feature values, e.g. histogram matching · CPC title

• G06K9/0014

  Physics · mapped topic

• G06T2207/20072

  Graph-based image processing · CPC title

• G06T2207/30096

  Tumor; Lesion · CPC title