HEP-2 CELLS CLASSIFICATION VIA FUSION OF MORPHOLOGICAL

AND TEXTURAL FEATURES

13 November 2012IEEE 12th International Conference on BioInformatics & BioEngineering

I. Theodorakopoulos, D. Kastaniotis, G. Economou and S. Fotopoulos

{iltheodorako, dkastaniotis}@upatras.gr , {economou, spiros}@physics.upatras.gr

Computer Vision GroupElectronics LaboratoryPhysics Department

University of Patras, Greece

www.upcv.upatras.gr www.ellab.physics.upatras.gr

Motivation

The standard screening test for detection of autoimmune diseases is the indirect immunofluorescence (IIF) test.

Human autoantibodies (AABs) associated with various autoimmune conditions, are detected by specific fluorescence patterns on a human epithelial cell line (HEp-2).

Testing is performed manually but : Requires highly-specialized personnel. Time-consuming procedure. Low standardization leads to high inter-laboratory

variance.

Typical IIF Procedure

Image Acquisition

Image Segmentation

Mitosis Detection

Fluorescence Intensity

Classification

Staining Pattern Recognition

Typical IIF Procedure

Image Acquisition

Image Segmentation

Mitosis Detection

Fluorescence Intensity

Classification

Staining Pattern Recognition

Input: Single-Cell Images Cell Contour Fluorescence

Intensity

Taxonomy

More than thirty different nuclear and cytoplasm patterns could be identified.

Can be grouped into six basic patterns:

Properties

Different staining patterns present variations both in: Morphological characteristics

Shape complexity Holes Intensity peaks

Textural characteristics Smooth Areas Grainy Areas

In order to capture the unique properties of each pattern, incorporation of both morphological and textural descriptors seems reasonable.

Multi-level Thresholding

Multi-level Thresholding

Multi-level Thresholding

Multi-level Thresholding

Multi-level Thresholding

Morphological Features

Cell’s contour complexity Threshold cell image into 9 levels equally spaced

between intensity extremes. Perform Connected Components Analysis on each

binary image. Discard blobs with area <1% of the mean area. On each binary image, compute:

Number of detected blobs Density of detected blobs Mean solidity of the detected blobs

Concatenate all features to a 28-dimensional descriptor

Local Binary Patterns (LBPs)

88 102

133

14 100

200

40 110 92

0 1 1 1 0 1 0 0

0 1 1

0 1

0 1 0

116

A well-established textural descriptor.

The biggest part of textural information is encoded in the 58 uniform patterns.

LBPs are not rotation invariant. A simple solution is to calculate the uniform

LPBs histograms on 80 rotated instances of the cell image (4.5 deg intervals).

Local Binary Patterns (LBPs)

Classification

The 28 morphological features and the 58-bin LBPs descriptor are concatenated in a 86-Dimensional feature vector.

Classification is performed using non-linear SVMs with Gaussian kernel.

Evaluation

Lack of publicly available datasets.

Evaluation on the dataset of HEp-2 cell classification contest (hosted by ICPR 2012 conference) 721 single-cell fluorescence images. Manually segmented and annotated by specialists in order to

provide ground truth. Binary masks and fluorescence intensity are provided. Approximately uniform distribution of patterns.

There are not reported results for comparison yet.

Results

K-fold validation - Classification performance of the various feature

sets for variable k

Confusion Matrix for 10-fold validation procedure using morphological and

textural features’ fusion

Thank You

This research has been co-financed by the European Union (European Social Fund – ESF) and Greek national funds through the Operational Program "Education and Lifelong Learning" of the National Strategic Reference Framework (NSRF) - Research Funding Program: Heracleitus II. Investing in knowledge society through the European Social Fund.

Acknowledgment