2018 IEEE Canadian Conference on Electrical & Computer Engineering (CCECE)

Diabetic Retinopathy Detection Using Machine

Learning and Texture Features
Mohamed Chetoui, Moulay A. Akhloufi, Mustapha Kardouchi
Perception, Robotics, and Intelligent Machines Research Group (PRIME),
Dept. of Computer Science, Université de Moncton
Moncton, NB, Canada
{emc7409, moulay.akhloufi, mustapha.kardouchi}@umoncton.ca

Abstract—Diabetic retinopathy (DR) is a medical condition 4]. Figure 1 shows examples of a normal retina and a DR
due to diabetes mellitus that can damage the patient retina and damaged retina.
cause blood leaks. This condition can cause different symptoms
from mild vision problems to complete blindness if it is not timely
treated. Hemorrhages, hard Exudates, and Micro-aneurysms
(HEM) that appear in the retina are the early signs of DR. Early
diagnosis of HEM is crucial to prevent blindness. Textures
features such as LBP have been widely used in the past as a
technique for DR detection. In this work, we introduce the use of
different texture features for DR, mainly Local Ternary Pattern
(LTP) and Local Energy-based Shape Histogram (LESH). We
show that they outperform LBP extracted features. Support Fig. 1. Example of a normal retina (left) and diabetic retinopathy (right)
Vector Machines (SVM) are used for the classification of the
extracted histogram. A histogram binning scheme for features Many researchers have devoted their effort to develop an
representation is proposed. The experimental results show that automatic Computer-Aided Detection (CAD) system for DR.
LESH is the best performing technique with an obtained accuracy Various approaches have been proposed [1-4], but HEM
of 0.904 using SVM with a Radial Basis Function kernel (SVM- detection in the retinal fundus images and its accuracy are still a
RBF). Similarly, the analysis of the ROC curve shows that LESH big challenge. The limitations are mainly due to the spherical
with SVM-RBF gives the best AUC (Area Under Curve) shape of the eye, leading to a brighter region in the center of the
performance with 0.931. retina and dark regions in the borders. Non-uniform lighting,
low contrast, small lesions, and the presence of objects in a
Keywords— Diabetic retinopathy, LESH, LTP, machine normal retina that have similar attributes to HEM [1] such as
learning, SVM. optic disc, blood vessels, etc. make it difficult to accurately
identify HEM. HEM is characterized by tiny spots that can be
I. INTRODUCTION lost during image processing.
In the last decade, the analysis of internal eye and retinal Some work has been proposed to detect DR using texture
fundus images has played a key role in human recognition features to preserve the HEM structure. In [2], the authors use
(biometrics), and in acquiring vital information about the health Local Binary Patterns (LBP) for the detection of hemorrhages
of the human visual system, thus helping in early detection of and exudates (HMA). The tests were performed on a database of
some pathologies such as glaucoma and diabetic retinopathy [1]. 89 images. They obtained an average accuracy of 86.15% and
an AUC of 0.87. In [3], the authors used fluorescein angiography
The eye is one of the most important and sensitive part of the
(FA) fundus images to detect micro-aneurysms (MA). They
human body. The inner part of the eye, facing the lens, includes
used Radon transform (RT) and multi-overlapping windows.
retina, optic disc, macula, and fovea. It can be seen during an
Tests were conducted on three databases of respectively 120, 50,
eye exam by looking through the pupil. The retina is sensitive to
and 22 images. The authors report the obtained best results for
light. Ophthalmologists examine retinal images manually. These
the two first databases with respectively sensitivity and
images, known and retinal fundus images, are captured by
specificity of 94% and 75% for the first database and 100% and
specialized cameras and are used in medical diagnosis.
70% for the second one. MA can be easily detected using this
Diabetic retinopathy (DR) is a condition where the retina of imaging modality, but the procedure needs the administration of
a diabetic patient is damaged due to blood leaks from retina some injections to the patient, making this approach less
blood vessels [1-4]. If not timely treated, the patient may suffer interesting as it can cause non-desirable health effects. The best
from complete blindness. DR progresses without showing any way to examine the retinal fundus is still to use standard retinal
symptoms in its early stages. To avoid serious complications, fundus imaging technologies. In [4], discrete wavelet transforms
patients are advised to make regular eye exams. The exudates are used. The wavelet decomposition was performed up to the
and micro-aneurysms are the first signs that appear in the retina second level. Eight energy features were extracted. Two features
at the initial stage of DR. The next stage is the appearance of from the coefficients of two levels and six energy values in three
hemorrhages. Early diagnosis of HEM at an early stage is crucial orientations (horizontal, vertical, and diagonal). Classification
in order to avoid severe complications leading to blindness [1- was performed using Support Vector Machines (SVM) with

978-1-5386-2410-4/18/$31.00 ©2018 IEEE

 2018 IEEE Canadian Conference on Electrical & Computer Engineering (CCECE)

different kernels. The authors collected 240 retinal fundus III. FEATURE EXTRACTION
images (120 normal and the remaining with different degrees of The proposed approach extracts texture features for learning
DR). They reported a result for accuracy, sensitivity, and and classification, more specifically: Local Ternary Pattern
specificity of 99% using SVM with polynomial kernel of order (LTP) and Local Energy-based Shape Histogram (LESH)
3. The results of these research works are interesting, however described in the following.
the number of images used for experimental tests is small. In
recent years, deep learning techniques were proposed for DR A. Local Ternary Pattern (LTP)
detection. Gulshan et al. [5] trained an Inception-v3
Convolutional Neural Networks (CNN) using a dataset of Local Ternary pattern texture descriptor was proposed in [8]
128,175 retinal images, they report an AUC performance of as an improvement to Local Binary Pattern descriptor [7]. Since
0.991. Ting et al. [6] trained a deep convolutional network in LBP is known to be sensitive to noise, LTP uses a fixed
71,896 retinal images and obtained an AUC of 0.936. They threshold to make binary patterns extraction more robust. LTP
experimented using 10 additional datasets, and obtained an computation is given by equation 1:
AUC ranging from 0.889 to 0.983. Deep learning techniques are 1 ≥
data hungry and need large image numbers for training. In this
, , =∑ ( − )× 3 , ( )= 0 | | <  
work we use an available public dataset that has a limited
number of images. −1 ≤ −
where τ is a threshold specified by the user, Pc is the central
Past work dealing with DR using texture features, has mainly pixel, Pi the neighboring pixel. To reduce the feature dimension,
focuses on Local Binary Patterns (LBP) [7] and wavelet the authors in [8] divide each ternary model into two parts (the
features. In this work, we propose the use of new texture features positive part and the negative part). Figure 2, shows the results
to detect DR, mainly Local Ternary Patterns (LTP) [8] and Local of performing LTP and LBP in a grayscale retinal fundus image.
Energy-based Shape Histogram (LESH) [9]. The extracted
features will be used to build a histogram that can be feed to an
SVM classifier [10]. A feature histogram binning is proposed to
reduce the number of features to be classified, while still getting
high performances. Different kernels are analyzed, and the best
performing techniques are presented in the experimental results.
The performance of the proposed techniques is evaluated using
MESSIDOR dataset [11] containing 1200 images. Average
accuracy (ACC) and area under ROC (Receiver Operating
Characteristic) curve (AUC) are measured in a k-fold cross-
validation scheme.
II. PROPOSED APPROACH AND DATABASE
For the extraction of HEM and the detection of DR, we Fig. 2. LTP (from left to right): 1st row: original image, and LBP; 2nd row: LTP
propose the use of texture features extracted from the retinal lower, and LTP upper
fundus images. SVM is used to learn the features characterizing
normal images from those showing the presence of DR. After the image processing using LTP, two feature
histograms are obtained for each image (upper and lower). Each
For experimental validation, we used a public database
histogram is normalized using a bin-width of 25. The results are
called MESSIDOR [11]. This database contains 1200 color
two vectors of dimension 25 each. In this work, the best
images of the eyes retinal fundus acquired by 3 ophthalmologic
performing threshold was equal to 5 [12]. A similar histogram
departments. They used a 3 CCD camera on a Topcon TRC
was extracted from the LBP image and can serve as a baseline
NW6 non-mydriatic retinograph with a 45 degrees field of view.
for comparison with other LBP texture-based DR techniques.
The captured images have resolutions of 1440*960, 2240*1488
or 2304*1536 pixels and they are encoded using 8-bits per Preliminary results, showed that the LTP upper was
channel. Medical experts provided the diagnosis for each image performing better than the lower counterpart. In the remaining
and evaluated the presence of DR and a retinopathy grade. 540 we will use the upper histogram.
images were graded as normal, 153 as mild, 247 as moderate,
and 260 as severe. B. Local Energy model and Local Energy-based Shape
Histogram (LESH)
In this work, we want to detect if DR is present or not, only
two classes will be considered. The database is labeled as normal The local energy model developed in [13] collects the
for all the images with no DR and as abnormal for all the images characteristics at points of an image where the local frequency
with DR, no matter the grade of the retinopathy. In addition, components are maximized in term of phase. To obtain the
since we are interested in extracting structural features using the phase, the image is convolved with a bank of Gabor filters with
retinal image fundus, we convert the color images to grayscale 5 spatial frequencies and 8 orientations. At each image location,
for further processing. for each scale and direction, it computes the magnitude and
phase of the associated pixel. Originally in [12], the authors
proposed to use the cosine of the difference of each phase
 2018 IEEE Canadian Conference on Electrical & Computer Engineering (CCECE)

component of the average phase as a measure of the phase Different kernel functions are available such as Radial Basis
symmetry. However, this measure is sensitive to noise and Functions (RBF), polynomial, etc.
results in a poor location. The framework was extended in [14]
to use a modified measure consisting of the phase deviation sine, V. EVALUATION METRICS
including a weighting factor of the frequency and a noise In medical imaging and diagnosis, detection methods are
cancellation factor. The proposed measure is invariant to non- evaluated using statistical measures such as sensitivity,
uniform lighting and contrast variation present in the image. specificity, accuracy, and Receiver Operating Characteristic
This local energy model detects the points of interest in the curves (ROC) [15]. Accuracy and Area under the ROC curve
images with high reliability even in the presence of light (AUC) are the most popular and will be used in this work.
variation and noise. In [14], the author proposes the
measurement of normalized energy, also called Phase The accuracy (ACC) is defined by equation 3 as follows:
congruency. = ( + )⁄( + + + ) (3)
The local orientation energy response varies depending on Where TP = number of true positives, TN = number of true
the underlying form such as corners, edges, and boundaries. We negatives, FN = number of false negatives, and FP = number of
compute a local histogram accumulating the local energy along false positives. The accuracy describes the overall rate of the
each Gabor filter orientation on different sub-regions of the retinal fundus images that were classified correctly.
image. The local histograms are taken from different sub-
Another measure used in this work is the Area under the
regions of the image, and then concatenated together. We do this
curve (AUC). It computed as the total area under the ROC curve
by getting a map orientation label where each pixel is assigned
and measures of how well we can distinguish between two
the label of direction where it has the largest energy. The local
diagnostic classes (abnormal/normal). AUC is particularly
histogram h is computed using equation 2:
useful in the presence of unbalanced data (large number of
ℎ , = ∑ × × ( − ) normal vs disease).
Where the subscript b represents the current box, L is the map of VI. EXPERIMENTAL RESULTS
orientation label, E is the local energy as proposed in [9], and w
As mentioned above, the MESSIDOR database images were
is a Gaussian weighting centered on region r. This weight is used
organized into two classes labeled as 0 (non-DR) and 1 (DR).
to provide soft margins through the bins by allowing small
LTP and LESH descriptors were extracted from the images
weighted overlap between neighboring sub-regions to overcome
(figure 4) and used for learning and classification. The data was
scale variation problems. In our experiments, for a 32x32 area,
divided into 10 sets. A 10-fold cross validation was performed
the Gaussian standard deviation σ is set to 20 in both directions.
on these sets and average results were computed.
The resulting feature histogram is a vector of 512
dimensions. Figure 3, shows the results of preprocessing a
grayscale retinal fundus image and the resulting Local energy
map and orientation map.

Fig. 3. LESH (image of figure 2) Local Energy map, and Orientation map
Fig. 4. Example of preprocessed images (from left to right): LTP upper, Local
IV. CLASSIFICATION Energy map, and Orientation map
SVM is used for learning and classification of the extracted
features. The database is divided into two classes of normal Many tests were conducted to select the best SVM kernels
retinal images and DR images. and optimize the hyper-parameters. SVM with RBF was the best
performing followed by the polynomial kernel for the LTP
The feature vectors of LTP and LESH are fed to SVM in accuracy.
order to learn from the extracted vectors the patterns that best
represent the retinal fundus images of a normal eye vs an eye Table I gives the ACC (accuracy) and AUC (Area under the
showing signs of DR. The learned models can be used later for ROC curve) results for the different tests. We can see that LTP
the classification of the images into DR or Non-DR classes. gives a very interesting performance. However, LESH is the best
performing and obtain an ACC of 0.904 using SVM with RBF
Basically, SVM is a linear classifier that extracts the vectors kernels. The same higher performance is obtained by LESH for
defining the best margin that separate different classes of data. AUC with RBF kernels which gives 0.931. These results
When the data in not linearly separable, kernel tricks are used to outperform LTP counterparts and show the robustness of LESH
define a kernel function that help project feature vectors into a features in classifying DR. In the other hand, we can see that the
higher dimensional space where they become linearly separable. LBP performance is lower than LTP and LESH.
 2018 IEEE Canadian Conference on Electrical & Computer Engineering (CCECE)

TABLE I. CLASSIFICATION RESULTS The proposed approach is suitable even for small datasets.
New techniques based on deep learning are data hungry but
ACC AUC
show impressive performances in different classification tasks
SVM Lin Poly RBF Lin Poly RBF including DR. Future work includes benchmarking the
performance of deep learning techniques and the proposed
LBP 0.805 0.813 0.823 0.847 0.884 0.897 texture-based features in a small dataset such as MESSIDOR.
LTP 0.822 0.869 0.841 0.881 0.914 0.916
REFERENCES
LESH 0.865 0.881 0.904 0.904 0.917 0.931
[1] S. Mohammadian, A. Karsaz and Y. M. Roshan, "A comparative analysis
of classification algorithms in diabetic retinopathy screening," 2017 7th
Figure 5 shows the ROC curves used for the computation of International Conference on Computer and Knowledge Engineering
AUC for the best obtained performance of each texture analysis (ICCKE), Mashhad, pp. 84-89, 2017.
technique and for both measures (shown in bold in Table I [2] M. N. Ashraf, Z. Habib and M. Hussain, "Texture Feature Analysis of
above). We can clearly see that LESH with SVM-RBF quickly Digital Fundus Images for Early Detection of Diabetic Retinopathy,"
converges to the best results and obtains more true positives for 2014 11th International Conference on Computer Graphics, Imaging and
Visualization, Singapore, pp. 57-62, 2014.
a small number of false positives compared to LTP and LBP.
This explains the obtained performance of LESH in term of both [3] M. Tavakoli, R. Shahri, H. Pourreza, A. Mehdizadeh, T. Banaee and M.
Bahreini Toosi, "A complementary method for automated detection of
ACC and AUC. microaneurysms in fluorescein angiography fundus images to assess
diabetic retinopathy," Pattern Recognition, vol. 46, no. 10, pp. 2740-
2753, 2013.
[4] K. Noronha, U. R. Acharya, K. P. Nayak, S. Kamath, and S. V. Bhandary,
"Decision support system for diabetic retinopathy using discrete wavelet
transform," Proc Inst Mech Eng H, vol. 227, no. 3, pp. 251–261, 2013.
[5] V. Gulshan, L. Peng, M. Coram, M. Stumpe, D. Wu, A. Narayanaswamy,
S. Venugopalan, K. Widner, T. Madams, J. Cuadros, R. Kim, R. Raman,
P. Nelson, J. Mega and D. Webster, "Development and Validation of a
Deep Learning Algorithm for Detection of Diabetic Retinopathy in
Retinal Fundus Photographs," JAMA, vol. 316, no. 22, pp. 2402–2410,
2016.
[6] D. Ting, C. Cheung, G. Lim, G. Tan, N. Quang, A. Gan, H. Hamzah, R.
Garcia-Franco, I. San Yeo, S. Lee, E. Wong, C. Sabanayagam, M.
Baskaran, F. Ibrahim, N. Tan, E. Finkelstein, E. Lamoureux, I. Wong, N.
Bressler, S. Sivaprasad, R. Varma, J. Jonas, M. He, C. Cheng, G. Cheung,
T. Aung, W. Hsu, M. Lee and T. Wong, "Development and Validation of
a Deep Learning System for Diabetic Retinopathy and Related Eye
Fig. 5. ROC curve for the best perfroming LESH, LTP and LBP Diseases Using Retinal Images From Multiethnic Populations With
Diabetes," JAMA, vol. 318, no. 22, pp. 2211–2223, 2017.
VII. CONCLUSION [7] T. Ojala, M. Pietikäinen, and D. Harwood,"A comparative study of
texture measures with classification based on featured distributions,"
The increase in diabetes cases limits the ability of current Pattern Recognition, vol. 29, no. 1, pp. 51–59, 1996.
manual testing. New algorithms for assisted diagnosis are [8] X. Tan and B. Triggs, "Enhanced Local Texture Feature Sets for Face
becoming very important today. Early detection of diabetes can Recognition Under Difficult Lighting Conditions," in IEEE Transactions
help the patients and limits the bad health consequences such as on Image Processing, vol. 19, no. 6, pp. 1635-1650, 2010.
blindness. Using retinal fundus images can help automate the [9] S. K. Wajid and A. Hussain, "Local energy-based shape histogram feature
diagnosis. Micro hemorrhages and aneurysms, known as HEM, extraction technique for breast cancer diagnosis," Expert Systems with
Applications, vol. 42, no. 20, pp. 6990–6999, 2015.
are the early signs of diabetic retinopathy (DR) and are difficult
to identify because of their similarities with normal parts of a [10] A. Ben-Hur, C. S. Ong, S. Sonnenburg, B. Schölkopf, and G. Rätsch,
"Support Vector Machines and Kernels for Computational Biology,"
healthy human. Other problems such as non-uniform lighting, PLOS Computational Biology, vol. 4, no. 10, p. e1000173, 2008.
low contrast, etc. can lead to a bad diagnosis. Texture based [11] E. Decencière, X. Zhang, G. Cazuguel, B. Lay, B. Cochener, C. Trone, P.
techniques for DR detection were proposed in the past. Most of Gain, R. Ordonez, P. Massin, A. Erginay, B. Charton, and J-C. Klein.
these techniques use LBP and wavelets for feature extraction "Feedback on a publicly distributed database: the Messidor database,"
from retinal images. In this work, we propose the use of new Image Analysis & Stereology, v. 33, n. 3, p. 231-234, 2014.
texture features, mainly LTP and LESH. These techniques [12] A.M. Owen, J.L. Iddon, J.R. Hodges, B.A. Summers, T.W. Robbins,
capture the local relationship between neighboring pixels and "Spatial and non-spatial working memory at different stages of
Parkinson's disease," Neuropsychologia, Volume 35, Issue 4, pp. 519-
features and are less sensitive to variation in illumination, color, 532, 1997.
noise, etc. These features extracted from the retinal fundus [13] R. Owens, S. Venkatesh, and J. Ross, "Edge detection is a projection,"
images are used to learn signs of HEM and differentiate between Pattern Recognition Letters, vol. 9, no. 4, pp. 233–244, 1989.
DR and non-DR. SVM is used to classify these features. [14] P. Kovesi, "Phase Congruency Detects Corners and Edges," Proc. VIIth
Polynomial and RBF kernels were the best performing in the Digital Image Computing: Techniques and Applications, Sydney,
detection of DR. Experimental tests with the MESSIDOR Australia, pp309-318, 2003.
database show that LESH and LTP perform well in detecting [15] J. A. Hanley and B. J. McNeil, ‘The meaning and use of the area under a
DR. Overall, LESH was the best performing with an accuracy receiver operating characteristic (ROC) curve’, Radiology, vol. 143, no.
of 0.904 and an AUC of 0.931 using SVM with RBF kernels. 1, pp. 29–36, 1982.