Abstract NVGE Voorjaarscongres 2013

jamaicacooperativeΤεχνίτη Νοημοσύνη και Ρομποτική

17 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

83 εμφανίσεις

TITLE

(
108 / 300 CHARACTERS

USED)

COMPUTER
-
AIDED

DE
LINEATION

OF EARLY NEOPLASIA IN
BARRETT’S

ESOPHAGUS USING
HIGH DEFINITION ENDOSCOPIC IMAGES

AUTHORS
(NO LIMIT)

F. van der Sommen
1
, S. Zinger
1
, P. de With
1
, E. J.
Schoon
2

1)
Electrical Engineering, Eindhoven University of Technology, Eindhoven, Netherlands

2)
Gastroenterology and Hepatology, Catharina Hospital, Eindhoven, Netherlands

ABSTRACT

(
2.6
87

/ 2.700 CHARACTERS U
SED
)

A
denocarcinoma of the esophagus is the faste
st rising type of cancer in the Western world. The recent
development of High Definition (HD) endoscopy has enabled the specialist physician to identify Barrett's
cancer at an early stage. Nevertheless, it still requires considerable effort, training and e
xpertise to be
able to recognize these irregularities associated with early cancer.

As a first step towards a Computer
-
Aided Detection (CAD) system that aids the physician in finding these
early stages of cancer, we
propose

an algorithm that is able to
aut
omatically detect early Barrett’s cancer,
using HD endoscopic images. Since early Barrett’s cancer is characterized by different color and texture
patterns, we employ features for capturing color and texture information. For our system, we have
selected

th
e Color Histogram (CH) and Gabor features and we compare with other well
-
kn
own color and
texture features
. The
system

employs tile
-
based processing,
which enables the delineation of regions
containing early cancer.
For classification

of these tiles
, we emp
loy
the widely used

Support Vector
Machine (SVM) and evaluate its performance using different parameters and kernel functions.


For clinical evaluation of the proposed algorithm, we
selected a total of 103 images of 30 patients

and let
an expert physician
delineate the
early cancer. We split the images into square tiles of different
dimensions, ranging from 25
×
2
5

to
250×250

pixels. Using the delineations of the expert, we split the tiles
into two classes, one with the tiles containing early cancer and another one with the tiles containing non
-
dysplastic Barrett’s tissue. This was done for each tile size separately. We used t
he tiles of 21 patients for
training and testing a classifier, where we ensured that no patient was in both the training and the test set.
The images of the other 9 patients were used for evaluation of the automatic delineations.

T
able 1 shows the classifi
cation performance of our classifier on the test sets

on tiles of
50×50

pixels
.

This tile size provides the best results in general.

The
proposed
system achieves a classification accuracy
9
4.2
%
on these tiles
of normal and of tumorous tissue and reaches an

Area Under the Curve (AUC) of
0.
986
.
Figure

1

show
s

the delineations of an expert gastroenterologist versus the delineations made by
the algorithm.

Our experiments and clinical validation of the results show that our approach is promising for a computer
-
a
ided detection system that helps the endoscopist in finding early stage Barrett's cancer. Further research
and clinical validation
i
s
needed

for the development of a system for real
-
time video analysis of
endoscopic images.

Table
1

Classification accuracy (%) on 50×50 pixel tiles : Proposed approach (GAB) vs. popular alternatives over different color spac
es

Color space
\

method

GAB

HOG

TSH

LBP

HOM

CON

ENE

ENT

DIS

COR

RGB

94.2

88.2

82.2

82.8

82.5

85.2

87.4

84.4

84.3

87.7

YCbCr

92.5

84.0

86.5

81.8

82.8

89.0

92.3

89.0

86.2

82.8

HSI

89.9

81.7

78.7

79.7

86.8

79.1

88.1

88.7

82.8

77.0

Abstract NVGE Voorjaarscongres 2013


Figure
1

Delineations of an expert gastroenterologist
(top
-

yellow)
and delineations of the proposed algorithm
(bottom
-

green)