4. Introduction
Figure 2: Diabetic macula edema
(swelling of the retina)
Diabetic retinopathy occurs when elevated blood sugar
levels cause blood vessels in the eye to swell and leak
into the retina.
4
5. Introduction
Abnormalities of Diabetic Retinopathy
•
•
•
•
Microaneurysms
Hemorphages
Cotton wool spots ( Soft Exudates)
Hard Exudates
Aim of this research is to develop system for detection
of hard exudates in diabetic retinopathy using nondilated diabetic retinopathy images
5
7. Methodology
Phase 1
Phase 2
Mathematical Morphology
Fuzzy Logic
• Exudates are identified
using mathematical
morphology
• Identified exudates are
classified as hard exudates
using fuzzy logic
7
9. Preprocessing
Input
Fundus Image
• Fundus Image is
performed by
fundus camera
Step 1
Step 2
Step 3
Step 4
Color Space
Conversion
Median
Filtering
Contrast
Enhancement
Gaussian
Filtering
• RGB color space
in the image in
converted to HIS
space
• Noise
suppression
• Contrast limited
adaptive
histogram
equalization was
applied for
contrast
enhancement
• Noise
Suppression
further
9
10. Optic Disc Elimination
Input
Preprocessed
Image
• Output of
preprocessing
stage
Step 1
Closing
• Closing operator
with flat disc
shape
structuring
element is
applied
Step 2
Step 3
Step 4
Thresholding
Large
Connected
component
Optic disc
elimination
• Image is
binarized
• P-tile method
and nilblack’s
method
• Connect all
regions
10
11. Exudates Detection
• Optic disc
eliminated
Image
• Standard
Deviation
• Remove optic
disc boundary
• Marker Image
• Difference
Image
I
n
p
u
t
• Closing
• Thresholding
• Fill holes
• Morphological
Reconstruction
• Result is
superimposed
11
15. Membership function of XB
Membership
function name
Parameters
[sig1 c1 sig2 c2]
B1
[0.217 0 3.081 5.408]
B2
[3 17 3 50]
B3
[3 60 3 102]
B4
[3 112 3 255]
Gaussian combination membership function
15
16. Membership function of Xout
Membership
function name
NotHardExudate
Parameters
[sig1 c1 sig2 c2]
[0.0008493 0 0.06795 0.07]
weakHardExudate [0.03 0.35 0.03 0.55]
mediumHardExud
ate
[0.03 0.65 0.03 0.75]
hardExudate
[0.03 0.85 0.03 0.9]
severeHardExudat [0.0161 0.9733 0.0256 1]
e
Gaussian combination membership function
16
17. Fuzzy rules
1
If (Xr is R1) Or (Xg is G1) Or (Xb is B4) Then (Xout is notHardExudate)
2
If (Xr is R2) And (Xg is G2) Or (Xb is B1) Then (Xout is weakHardExudate)
3
If (Xr is R2) And (Xg is Not G2) And (Xb is Not B1) Then (Xout is notHardExudate)
4
If (Xr is R3) And (Xg is G3) And ((Xb is B1) Or (Xb is B2) ) Then (Xout is weakHardExudate)
5
If (Xr is R3) And (Xg is G3) And (Xb is B3) Then (Xout is notHardExudate)
6
If (Xr is R3) And (Xg is Not G3) Then (Xout is notHardExudate)
7
If (Xr is R4) And (Xg is G3) And (Xb is B1) Then (Xout is mediumHardExudate)
8
If (Xr is R4) And (Xg is G3) And (Xb is B2) Then (Xout is weakHardExudate)
9
If (Xr is R4) And (Xg is Not G3) Then (Xout is notHardExudate)
10
If (Xr is R5) And ((Xg is G2) Or (Xg is G3) Or (Xg is G4)) Then (Xout is notHardExudate)
11
If (Xr is R5) And (Xg is G5) And ((Xb is B1) Or (Xb is B2)) Then (Xout is HardExudate)
12
If (Xr is R5) And ((Xg is G6) Or (Xg is G7)) Then (Xout is notHardExudate)
13
If (Xr is R5) And (Xb is B3) Then (Xout is notHardExudate)
14
If (Xr is R6) And ((Xg is G2) Or (Xg is G3)) Then (Xout is notHardExudate)
15
If (Xr is R6) And (Xg is G4) And ((Xb is B1) Or (Xb is B2)) Then (Xout is HardExudate)
16
If (Xr is R6) And (Xg is G5) And ((Xb is B1) Or (Xb is B2)) Then (Xout is HardExudate)
17
If (Xr is R6) And (Xg is G6) And ((Xb is B1) Or (Xb is B2)) Then (Xout is HardExudate)
18
If (Xr is R6) And (Xg is G7) Then (Xout is notHardExudate)
19
If (Xr is R6) And (Xb is B3) Then (Xout is notHardExudate)
20
If (Xr is R7) And (Xg is G6) And ((Xb is B1) Or (Xb is B2) Or (Xb is B3)) Then (Xout is severeHardExudate)
21
If (Xr is R7) And (Xg is G5) And ((Xb is B1) Or (Xb is B2)) Then (Xout is notHardExudate)
22
If (Xr is R7) And ((Xg is G2) Or (Xg is G3) Or (Xg is G4)) Then (Xout is notHardExudate)
17
18. Implementation
• 38 images were used to testing
• Images were taken from Kuopio university
hospital
• The images’ size were 1500 , 1152 pixels
Tested using MATLAB 7.10
18
20. Results – Optic Disc Elimination
(a)
(b)
(c)
(d)
(e)
(f)
(a)-Applying morphological closing operator, (b)-Thresholded image using
Nilblack’s method, (c)– Thresholded Image using percentile method,
(d)- Large circular connected component, (e)-Inverted binary image,
(f)- Optic disc is eliminated from the preprocessed image
20
21. Results – Exudates Detection
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
(a)- Applying morphological closing operator , (b)- Standard deviation of the image ,
(c)-Thresholded image using triangle method , (d)- Unwanted borders were removed ,
(e)- Holes are flood filled , (f)- Marker Image , (g)- Morphological reconstructed image21
,
(h)- Thresholded image , (i)- Result is super imposed on original image
22. Results – Classification of Exudates
(a)
(b)
(c)
Performance
•
•
•
•
Overall sensitivity-81.76%
Specificity – 99.96%
Precision – 81%
Accuracy – 99.84%
(a)- Not exist diabetic retinopathy,
(b)- 42% of diabetic retinopathy hard exudates ,
(c)- 89% of diabetic retinopathy hard exudates ,
22
23. Future Works
•
•
•
•
Preprocessing Stage
Optic Disc Elimination
Exudates Detection
Classification of Exudates as Hard
Exudates
• Exudative Maculopathy Detection
• Support Vector Machines, K Means
Algorithms, Radial Basis Functions
Tested using MATLAB 7.10
23
26. References
• Meysam Tavakoli, Reza Pourreza Shahri, Hamidreza Pourreza, Alireza Mehdizadeh,
Touka Banaee, Mohammad Hosein Bahreini Toosi, A complementary method for
automated detection of microaneurysms in fluorescein angiography fundus images
to assess diabetic retinopathy, Pattern Recognition, Volume 46, Issue 10, October
2013, Pages 2740-2753, ISSN 0031-3203,
http://dx.doi.org/10.1016/j.patcog.2013.03.011.
(http://www.sciencedirect.com/science/article/pii/S0031320313001404)
• M. Usman Akram, Shehzad Khalid, Shoab A. Khan, Identification and classification
of microaneurysms for early detection of diabetic retinopathy, Pattern Recognition,
Volume 46, Issue 1, January 2013, Pages 107-116, ISSN 0031-3203,
http://dx.doi.org/10.1016/j.patcog.2012.07.002.
(http://www.sciencedirect.com/science/article/pii/S003132031200297X)
• R.H.N.G. Ranamuka, Automatic detection of diabetic retinopathy hard exudates
using mathematical morphology methods and fuzzy logic, Graduation Thesis,
University of Sri Jayewardenepura, 2011
26
Microaneursms is the early stage of Diabetic Retinopathy
Akara’s suggested certain steps for optic disc detection prior to the exudates identification.After the optic disc elimination mathematical morphology has been used for the exudates detectionAkara proposed a Fuzzy C Means (FCM) clustering method for exudates detectionSuggested a computer based approach for automated classification of Normal, NPDR and DPRThey have used the green layer for exudates detection because they have discovered that the brightness area including exudates of retinal image is in green layer in literature
Intensity band of the HIS image is used at this stageFirstly RGB color space in the original fundus image is converted to HIS (HUE, Intensity and saturation) space.Then median filter is applied for the intensity band of the image for the noise suppression. Median filter is non linear median filter which is used to remove noises in an image with minimal degradation to edges.Subsequently the Contrast limited adaptive histogram equalization was applied for contrast enhancement this adaptive histogram method is used to improve the local contrast of an image. It may be produce a significant noiseGaussian function is applied for noise suppression furtherThis gaussian filtering function is used to filter out the noise in the image without compromising on the region of interest
Firstly the closing operator with a flat disk shape structuring element is applied for the preprocessed image.Then the result image is binarized using thresholdingtechniqueClosing is a morphological operation
Remove optic disc boundaryTriangle method is used to obtain thesholded imageFill HolesMarker image The intensity band of original image is selected as the mask image.Morphological ReconstructionDifference Image
I have used the RGB color space values of retinal image to form the fuzzy set and the membership functionsNot Hard ExudatesWeak Hard ExudatesMedium Hard ExudatesHard ExudatesSevere Hard Exduates
There are 7 linguistic variables for Xr
There are 7 linguistic variables for Xg
There are 4 linguistic variables for Xr
There are 5 linguistic variables for output value
Fuzzy rules
Those 38 images were publicly available diabetic retinopathy dataset
Those 38 images were publicly available diabetic retinopathy dataset
Proposed Radon Transform method to detect MAs
ProposedHybrid Classifier which combines Gaussian Mixture Model and Support Vector Machine
Those 38 images were publicly available diabetic retinopathy dataset