The Multifocal ERG in Early Detection of Diabetic

Retinopathy
XU Jin
1
, HU Guangshu
1
, HUANG Tianna
2
, HUANG Houbin
2
CHEN Bin
2
1. Department of Biomedical Engineering, Tsinghua University, Beijing 100084, China
2. Department of Ophthalmology, General Hospital of Chinese PLA, Beijing 100853, China
Abstract- To assess retinal changes in diabetic eyes at early
stage of retinopathy, retinal responses were examined using
multifocal ERG. Seven healthy subjects, 16 diabetics with no
apparent retinopathy (NDR) and 9 diabetics who had back-
ground diabetic retinopathy (BDR) underwent multifocal
ERG testing. The first slice of the second order kernel (K21)
were extracted from multifocal ERG record and summed.
Three major peaks (P1, N1 and P2) on the trace of summed
K21 are evaluated across all subjects. With their amplitude
and implicit time as 6 features, linear classifiers were built.
These classifiers were used to discriminate eyes of control,
NDR and BDR subjects. The classification error dropped sig-
nificantly.
Keywords- diabetic retinopathy, early detection, multifocal
ERG, second order kernels
I. INTRODUCTION
Diabetic Retinopathy (DR) is a major cause of blind-
ness in modern society. It is a common microvascular com-
plication in patients with diabetes and may have a sudden
and debilitating impact on visual acuity, eventually leading
to blindness. DR is a progressive disease. According to se-
verity level, it includes the following stages: no apparent
DR, background DR, and proliferative DR [1]. To know
when to initiate treatment of diabetic retinopathy and to
prevent visual loss, testing methods are needed to detect
retinal changes as early as we can, better before BDR hap-
pened.
As hyperglycemia caused by diabetes impairs the ret-
ina from the beginning, most of retinal changes during this
pre-retinopathy period are functional and can not be ob-
served by traditional clinical methods, like ophthalmoscope.
Electrophysiological methods are looked as important tools
to find functional abnormities of retina in the early stage of
DR [2]. Multifocal electroretinogram (mfERG) is a newly
developed advanced electrophysiological technique [3]. It
enables simultaneous ERG testing of multiple small retinal
areas and allows a fast, objective evaluation of retinal func-
tion. Hood has made a comprehensive review on the
mechanism and applications of mfERG in 2001[4].
mfERG responses have shown significantly reduced
amplitudes and response delays in diabetic eyes, with or
without retinopathy [5,6,7]. The problem is that, in early
stages of DR, the infection of DR on mfERG response is
very weak [4]. So, recognition rate between healthy sub-
jects, subjects without DR and subjects with BDR is low at
present. For example: when amplitude of mfERG responses
was evaluated, only half of diabetic eyes were correctly
discriminated from normal eyes in [7]; in [6], 12 out of 18
diabetic eyes have the same range of implicit time with
those eyes of non-diabetics. In this study, our purpose is to
increase the classification performance.
According to extraction rule, one first order kernel and
several high order kernels can be extracted out of every
mfERG record. And second order kernels are suggested
more sensitive for the detection of early changes in retinal
function of diabetics than the commonly considered first
order kernel [5,7]. In this study, the first slice of second
order kernel (K21) of mfERG was used.
II. DATA PREPARATION
Twenty five non-insulin-dependent diabetic pa-
tients(56.113.1 years of age, duration of diabetes, 5~30
years) were examined. 16 of them had no apparent DR,
attributed to group NDR; the other 9 had background DR,
attributed to group BDR. The NDR and BDR eyes were
classified by ophthalmoscopic examination and fluorescein
angiography. The control group was consisted of 7 volun-
teers with healthy retina(52.314.3 years of age). Pupils
were dilated maximally. The t-test results showed no dif-
ference between the results obtained from subjects in whom
only one eye was analyzed and those obtained from sub-
jects in whom both eyes were analyzed. Therefore, results
from both eyes of every subject were included in this study.
MfERG tests were performed on a visual evoked re-
sponse imaging system (Veris Science
TM
5.2). The visual
stimulation consisted of 103 scaled hexagons array on a
CRT screen placed in front of subjects and stimulate the
central 50of the retina (Fig. 1). Each stimulus hexagons
flickers between two binary states (black or white) con-
trolled by a predetermined m-sequence. A shifted version of
the same sequence can be used to control different areas.
Retinal responses were derived from the cornea by means
of a Burian-Allen bipolar contact lens electrode, amplified
(50,000 gain) and band pass filtered 10300 Hz. By per-
forming a cross-correlation between the m-sequence for a
particular area and the mixed retinal response with a com-
puter, it is possible to recover the small independent multi-
focal ERG responses.
As shown in Fig.1, every mfERG recording produces
103 focal K21s corresponding with 103 regions on the ret-
ina. To simplify the comparison, all K21s were summed
across the entire array for each mfERG recording. (The
word K21 in following discussion means this summed re-
sult.)
Proceedings of the 2005 IEEE
Engineering in Medicine and Biology 27th Annual Conference
Shanghai, China, September 1-4, 2005
0-7803-8740-6/05/$20.00 2005 IEEE.
7762
Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on August 18, 2009 at 23:09 from IEEE Xplore. Restrictions apply.
Fig.1. The schematics of a multifocal ERG system.
III METHOD AND RESULT
A typical summed K21 is shown in Fig.2. It has a
waveform structure including three major peaks (P1, N1,
and P2).
Fig.2. Typical waveform of summed K21
Comparing K21 from control, NDR and BDR subjects,
the amplitude and implicit time of all three peaks changed
in different level. And these two features of peak P2 scat-
tered in a wider range and alter more significantly than that
of P1 and N1. The distribution of peak P2 among subjects
from three groups was shown in Fig.3. In Fig.3 (a), the im-
plicit time of P2 obviously distributed in a narrower area
within control group. And the amplitude of P2 reduced sig-
nificantly along with the progression of DR in Fig.3 (b).
The average amplitude is 2.50V for control group, 1.44V
for NDR group and 1.09V for BDR group. When the am-
plitude control range of P2 was set to 1.5V, the minimum
value among control subjects, 22 out of 32 NDR subjects
and 15 out of 18 BDR subjects fell below the range line.
This result is not good enough to discriminate eyes of sub-
jects in this study.
(a)
(b)
Fig.3. The implicit time and amplitude of P2 of the summed K21. Those
* points represent data from control subjects, o represent NDR,
represent BDR. Short transverse lines in (b) indicate average amplitude of
every group. The dashed line indicates the range of control.
To increase the recognition rate, we synthesize im-
plicit time and amplitude of peak P2. Every point in Fig.4
has these two features as the x-coordinate and the
y-coordinate respectively. Those dark lines in Fig. 4 served
as the decision boundary of the two-feature linear classifiers.
The classification results in Fig.4 were also shown by con-
fusion matrix in Table.1.
7763
Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on August 18, 2009 at 23:09 from IEEE Xplore. Restrictions apply.
Fig. 4 Scatter plot of P2 from control, NDR and BDR subjects. Dark lines
indicate the decision line of linear classifiers.
Table.1. Confusion matrices from Fig.4
a) Control subjects versus NDR subjects
Classification Labels
True Labels
Control NDR
Totals
Control 12 2 14
NDR 3 29 32
b) Control subjects versus BDR subjects
Classification Labels
True Labels
Control BDR
Totals
Control 13 1 14
BDR 2 16 18
c) NDR subjects versus BDR subjects
Classification Labels
True Labels
NDR BDR
Totals
NDR 16 16 32
BDR 5 13 18
According to Fig.4 and Table.1, the classification error
significantly decreased when linear classifiers based on
amplitude and implicit time of peak P2 were used. But,
classification result between NDR subjects and BDR sub-
jects is still far from good.
To increase the classification performance, both of
amplitude and implicit time of peaks P1, N1 and P2 were
used as 6 features of linear classifiers. The linear discrimi-
nant function g(x) is given in formula (1). Among subjects
from three group, there are three two-category linear classi-
fiers need to be built. Value of the weight vector W and the
threshold weight w
0
were computed by minimizing the er-
rors in the least square sense (the algorithm is described in
[8]), and listed in Table.2.
0
1 1_ ImplicitTime
2 1_ ImplicitTime
3 2 _ ImplicitTime
( )
4 1_ Amplitude
5 1_ Amplitude
1 2 _ Amplitude
w P
w N
w P
g x w
w P
w N
P
T
( (
( (
( (
( (
= - +
( (
( (
( (
( (

(1)
Table.2. Weight vector and threshold weight of the linear discriminant
function

1
:
2
Control : NDR Control : BDR NDR : BDR
w
0
-8.84 -0.0552 121.95
w
1
0.138 0.217 7.52
w
2
0.00285 -0.371 -10.36
w
3
0.0858 0.0633 -0.00547
w
4
0.0532 0.364 2.33
w
5
0.267 -0.114 -6.27
Then compute g(x) of formula (1) for all subjects in
control, NDR and BDR groups. The unit of amplitude is
microvolt, and that of implicit time is millisecond. Classify
all subjects according to the sign of g(x):
1
2
if ( ) 0,
if ( ) 0,
g x x
g x x
e
e
> e

< e

(2)
Because the number of features is 6, these three linear
classifiers can not be visualized as in Fig.4. Confusion ma-
trices of these classifiers are given in Table.3.
Table.3. Confusion matrices for those classifiers with 6 features
a) Control subjects versus NDR subjects
Classification Labels
True Labels
Control NDR
Totals
Control 12 2 14
NDR 3 29 32
b) Control subjects versus BDR subjects
Classification Labels
True Labels
Control BDR
Totals
Control 14 0 14
BDR 0 18 18
c) NDR subjects versus BDR subjects
Classification Labels
True Labels
NDR BDR
Totals
NDR 30 2 32
BDR 4 14 18
The classification performance is further increased
compared with previous classifiers with two features. At
this situation, the early detection of diabetic retinopathy
becomes possible.
Above results prove the efficiency of pattern recogni-
tion method in mfERG application of early detection of
diabetic retinopathy. And there are still works need to be
done: In above discussion, only two-category classifier was
used. According to clustering of subjects in this study,
multi-class classifier should be designed to deal with the
early detection of DR. On the other hand, the amplitude and
implicit time of P1, N1 and P2 may not be the best features
in the classification, which suggests that feature extraction
may be executed before the classification.
7764
Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on August 18, 2009 at 23:09 from IEEE Xplore. Restrictions apply.
REFERENCE
[1] SC Brailsford, R Davies, C Canning and PJ Roderick, Evaluating
screening policies for the early detection of retinopathy in patients
with non-insulin dependent diabetes, Health Care Management
Science. Vol.1, pp. 115124, 1998
[2] R Tzekov, GB Arden, The electroretinogram in diabetic retinopa-
thy, Survey of Ophthalmology, Vol. 44(1), pp. 53-60, 1999
[3] EE Sutter and D Tran, The field topography of ERG components in
man--I. The photopic luminance response, Vision Research, Vol. 32,
pp. 433-466, 1992
[4] DC Hood, Assessing fetinal function with the multifocal technique,
Progress in Retinal and Eye Research, Vol. 19(5), pp. 607-646, 2000
[5] AM Palmowski, EE Sutter, MA Bearse, W Fung, Mapping of retinal
function in diabetic retinopathy using the multifocal electroretino-
gram, Investigative Ophthalmology and Visual Science, Vol. 38, pp.
25862596, 1997
[6] B Fortune, ME Schneck, AJ Adams, Multifocal electroretinogram
delays reveal local retinal dysfunction in early diabetic retinopathy,
Investigative Ophthalmology and Visual Science, Vol. 40, pp.
26382651, 1999
[7] Y Shimada, Y Li, MA Bearse, EE Sutter, W Fung, Assessment of
early retinal changes in diabetes using a new multifocal ERG proto-
col, British Journal of Ophthalmology, Vol. 85, pp. 414419, 2001
[8] OD Richard, EH Peter, GS David, Pattern Classification (2nd Edi-
tion), Wiley-Intersciencepp.239-249, 2000
7765
Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on August 18, 2009 at 23:09 from IEEE Xplore. Restrictions apply.