Automatic Detection of Diabetic Retinopathy

1) This document discusses methods for the automatic detection of diabetic retinopathy through the classification of anatomical and pathological features in retinal images. 2) It evaluates techniques like template matching, regression analysis, and histogram matching for localizing the optic disc and presents a hybrid approach that uses vessel detection and a search window. 3) Experimental results on two public datasets show the hybrid approach can accurately detect the optic disc in 90% of images, while statistical texture measures help distinguish true exudates from false ones.

Uploaded by

mmanish91

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

43 views17 pages

Automatic Detection of Diabetic Retinopathy

Uploaded by

mmanish91

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 17

Automatic detection of

Diabetic retinopathy

1
2
Classification of retinal image
• Anatomical features- Optic disc, Blood vessels
and Fovea.
• Pathological features- Microaneurysm (tinny
blood spots), Exudates (lipid – proteins),
Hemorrhages (bigger blood spots).

3
Data Collection
• Two database are obtained
 DRIVE Database
STARE Database
• Both the databases represent wide variability,
noisy and low resolution.
• Focus is put on dealing with noisy and low
resolution images.

4
OD Localization
• Vulnerable to be detected as exudates.
• Not always circular and No standard position
• Assumptions
Brightest part
Lies at the convergence of BVs
BVs emanate near‐vertically form OD

5
Methods used for OD localization
• Template matching
• Regression analysis
• Histogram matching
• Hybrid approach (proposed method).

6
Template matching method [2]

• In this method we use a

number of retinal images
to create a template for
optic disc as shown in fig.
•For this we use a window
with the typical size of the
optic disc (80 × 80 pixels)
to extract
the optic disc of each
retinal image.
• Then average image of
this template image is
obtained.
•Then, an 80 × 80 pixels
window is moved through
retinal image.

Fig.2 Training Images for Localization of Optic Disk

7
Disadvantage of template matching
• This method is most accurate method for the
detection of OD. But it involves high
computational cost (i.e. 80*80=6400 pixel
comparison at each point).
• To reduce this computational cost we propose
histogram matching approach. It involves only
256 gray level comparison at each point.

8
Histogram matching approach

In this method we obtain histogram of each

channel (i.e. RGB) individually. On
observing it is noted that green channel
histogram contains maximum information
so we use this channel histogram for OD
detection.

Fig.3 Histograms for three channels: (a) red, (b) green, (c) blue.
9
1
c
(1  (ai b i ) 2 ) I = no of gray levels
i

Challenges ???
• Two visually different images can have same
histogram and in pathological images
exudates have similar properties as of OD and
hence same histogram.
• Histogram matching fails in case of
pathological images.
10
First method
• We know main blood
vessels emerge from OD.
So by identifying Main
blood vessels roughly we
can estimate.
• To avoid the chance of
false detection due to
exudates we use a search
window of size 15*15 near
maximum correlated value
obtained and finds a gray
value belongs to main Fig.4 (a) original Retinal from DRIVE database
vessel obtained earlier if (b) Smoothed image after median filter 50*50
exists stops pixels c) main blood vessels obtained after
Morphological operation on foreground
detection d) centre located retinal image.
11
Results

Fig.5 Optic disc center for retina images in STARE dataset.

In STARE database we apply this method on 20 images and obtain OD accurately in

18 images. We achieved accuracy of 90percent.

12
Fig.6 Retina image where the optic disk covered with black circle

13
For Exudates
• Using Statistical approach
L 1
Mean Z i P ( Z i )
i 0

S .D  2 (Z )
1
Smoothness1 
(1  S .D 2 )
L 1
Uniformity   (Zi )
P 2

i 0

Where Zi is a variable intensity at location I, P(Z) is the histogram of intensity level

L is number of levels.

14
Texture measure for Retinal image

Texture Parameter True Exudates False Exudates

Average 5.677 1.8016
Smoothness 0.0213 0.0070
Uniformity .9565 .9860

On observing above parameter we conclude that exudates have more average

intensity but they are less uniform in nature they have sharp edges.

15
Future work
• In future I would like to obtain anatomical
features in Pathological retinal images.
• I would like obtain pathological features (MA,
Hemorrhages, and Exudates automatically).
• I would like to classify image as Normal, or DR
effected image.

16
Work plan
Semester Work decided Work done Database used