See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/340691417

Classification of Pathological Signs for Diabetic Retinopathy Diagnosis using

Image Enhancement Technique and Convolution Neural Network

Conference Paper · November 2019

DOI: 10.1109/ICCSCE47578.2019.9068538

CITATIONS READS

5 54

6 authors, including:

Abdul Hafiz ABU Samah Fadzil Ahmad

Universiti Teknologi MARA Universiti Teknologi MARA
4 PUBLICATIONS 18 CITATIONS 34 PUBLICATIONS 453 CITATIONS

SEE PROFILE SEE PROFILE

Muhammad Khusairi Osman Abd Aziz Nor Azimah

Universiti Teknologi MARA Universiti Teknologi MARA
132 PUBLICATIONS 1,099 CITATIONS 15 PUBLICATIONS 81 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Abd Aziz Nor Azimah on 05 February 2022.

The user has requested enhancement of the downloaded file.

2019 9th IEEE International Conference on Control System, Computing and Engineering (ICCSCE), 29 Nov.–1 Dec. 2019, Penang, Malaysia

Classification of Pathological Signs for Diabetic Retinopathy

Diagnosis using Image Enhancement Technique and
Convolution Neural Network
Abdul Hafiz Abu Samah, Fadzil Noritawati Md Tahir Nor Azimah Abd. Aziz
Ahmad, Muhammad Khusairi Faculty of Electrical Engineering, Faculty of Medicine,
Osman, Mohaiyedin Idris Universiti Teknologi MARA, Department of Ophthalmology
Faculty of Electrical Engineering 40450 Shah Alam, Selangor, UiTM Kampus Sungai Buloh,
Universiti Teknologi MARA, Malaysia 47000 Sungai Buloh Selangor,
Cawangan Pulau Pinang, Malaysia
13500 Permatang Pauh,
Penang, Malaysia.
Email:
[email protected]

Abstract—Diagnosis of diabetic retinopathy (DR) involves Recently, the development of intelligent screening and
visual examination of retinal images by ophthalmologist to detection system of the DR using CNN shows an increasing
detect pathological signs such as exudate, haemorrhage (HEM) trend among researchers. CNN is a class of deep neural
and microaneurysm (MA). This process is conducted manually, networks learning algorithm most commonly use to
therefore it is time-consuming and subjected to human error. analyzing visual imagery. CNN involves very little pre-
This paper develops an automatic and intelligent machine processing as it learns image feature that in traditional
learning algorithm for the detection of diabetic retinopathy algorithms were hand-engineered. This independence from
(DR) in fundus image. It involves image enhancement and
human effort and prior knowledge in complex feature design
classification of pathological signs using convolution neural
is a major advantage.
network (CNN) for the DR pathological signs classification. In
the image enhancement process, high-pass filter and histogram
Motivated by the work of [8], this paper proposed an
equalization are applied to improve visual quality of fundus automatic and intelligent machine learning algorithm for the
images. A five layers CNN architecture is implemented to detection of diabetic retinopathy (DR) in fundus image.
classify the three pathological signs; exudate, HEM and MA. However, this work is different with [8] in image
Two dataset, DIARETDB1 and e-Ophtha are used to evaluate enhancement approach and network architecture.
the performance of the system. Simulation results using
enhanced DR images show significant improvement in II. CLINICAL FEATURES
classification accuracy compared to those images without In order to diagnose the DR disease, several clinical
enhancement for both datasets. features need to be identified by ophthalmologist from the
fundus images of the suspected DR patient. The features
Keywords- Diabetic retinopathy, convolutional neural network, include MA, HEM and exudate as shown in Fig. 1.
histogram equalization, high pass filter.

I. INTRODUCTION
Diabetes mellitus (DM) which is a non-communicable
disease directly affecting millions of people worldwide.
World Health Organization (WHO) estimates by 2030,
Malaysia would have 2.48 million people with DM [1]. DM
is caused by insufficient or resistance of insulin in the human
body [2]. This disease may lead to organ complications. One
of most common affected organs is the retina of the eye. The
specific disease is known as diabetic retinopathy (DR).
Prolong and unattended DR may lead to visual loss. Early
detection and treatment will save cost and prevention from
blindness [3].
A traditional DR screening is through manual inspection
of the retina based non-digital ophthalmoscopy method. Fig. 1. Example of common pathological signs of DR.
However, this conventional method is time-consuming,
tedious, and may lead to inaccurate diagnosis [2]. In the past
three decades, an automatic screening of diabetic retinopathy MA is a formation of small, balloon-round shape and dark
using computer vision and intelligent image processing red spot occur due to high blood sugar (hyperglycemia) in
technique become popular among medical researchers. It help blood vessels walls [11]. The number on MA increase as
lessen the work-load of ophthalmologist and provide timely increase stage of DR. Dimension range of MA is between
diagnosis for its early treatment [2]. For automatic screening, 20um and 200um equivalent 1-3 pixel in different image [12].
several machine learning has been developed to detect HEM occurs due to leakages or ruptured of blood from
common DR pathological sign such as microaneurysm (MA), capillaries or blood vessel. It appears as red flamed, dot and
haemorrhages (HEM), and exudate [4]. These include the use blot in size [13]. Usually HEM is 125µm greater than MA in
of artificial neural network (ANN) [5], support vector size [14]. Exudate appears in bright white or yellowish-white
machine (SVM) [6], k-nearest neighbor (K-NN) [7] and colour with sharp edges against the background from blood
convolution neural network (CNN) [4], [8]–[10]. vessel [12]. It is due to lipoproteins leaking out from retinal
978-1-7281-3500-7/19/$31.00 ©2019 IEEE 221
2019 9th IEEE International Conference on Control System, Computing and Engineering (ICCSCE), 29 Nov.–1 Dec. 2019, Penang, Malaysia

vessel at outer layer of the retina. Exudate often arranged in

form of circinate shape rings [14]. Applying high-pass filter will enhance region of high intensity
gradient in the images, but significantly affect the colour
III. METHODOLOGY intensities of the images. Therefore, histogram equalization
(HE) was applied to enhance image and improve visual
A. Data Collection quality of the pathological signs. In fact, image enhancement
using HE technique is widely used in many applications
In this research, two public datasets were used which are
related to medical imaging [18]–[20].
DIARETDB1 [15] and e-Ophtha [16].
DIARETDB1: It consists of 89 colour fundus images of C. Patch split and labeling
which 84 contain at least mild non-proliferative signs After images enhancement process, the image was split
(Microaneurysms) of the diabetic retinopathy, and 5 images into patches with size of 50x50 pixel. The patch size is
are considered as normal which do not contain any signs of sufficient to cover the smallest pathological sign in the DR
the diabetic retinopathy according to all experts who images. All these patches were labelled into their associated
participated in the evaluation. Images were captured with groups.
unknown camera settings and 50 Field-Of-View (FOV). The
size and resolution all this photo using (1500 x 1152) pixel Patch images are labelled by comparing the original image
with (png format). with ground truth images corresponding to the three
pathological signs: exudate, HEM, MA and background (no
e-Ophtha: E-ophtha is made of two subset: (i)“e-ophtha- DR sign) as shown in Fig. 2. Fig. 3 shows example of patches
MA” which have contains 148 images with microaneurysms with different pathological signs and background. Image data
(MAs) or small haemorrhages, and (ii) “e-ophtha-EX “which augmentation was also applied on the patches to expand
has 47 images with exudates (EXs). In this database, image dataset in order to improve the ability of CNN to generalize.
size and resolution range from (1440 x 960) pixel to (2544 x
1696) pixel with jpg format. All fundus image was taking by
Topcon NW 100, Topcon NW 200 and Canon CR5-45NM.
This dataset are provided annotated ground truth or
reference for the sign MA and Exudate. This ground truth
images was be annotated by two specialists in each
ophthalmology department. Table I depicts the number of
normal and abnormal images in various databases.

TABLE I. NUMBER OF NORMAL AND ABNORMAL IN VARIOUS

DATABASE

Dataset No. of Image

Normal Abnormal
Fig. 2. Image follow the ground truth to identify position each sign category
before split.
DIARETDB1 5 84
e-Ophtha-MA 233 148
e-Ophtha-Ex 35 47

B. Image Enhancement
The one of problem retinal image has variety in term of
brightness and contrast which result hard to identify DR
feature. Moreover, every eye has their own unique shape and
feature [11]. There are several reason fundus images was not
able to take clearly from ophthalmologist such as pupil unable
to open (constrict) at bright light condition and aging cataract
clouding at eye lens which lead to less light unable to pass
through the pupil. Since colour of retinal image must be tuned
to allowed CNN performing robust, our proposed method to
enhance the image using high pass filter technique and
Fig. 3. Patches are grouped into MA, HEM, exudate and no-Sign.
histogram equalization to remove noise.
These resulted 5569, 3328, 4816 patches for MA, HEM,
High pass filter is a filter that removes low frequency
exudate and 6642 patches with no pathological sign for
(blur) signal to become sharp signal [17]. It is usually used for
DIARETDB1. While 2292, 4193 patches for MA, exudate and
edge detection and image sharpening. The transfer function of
6863 patches with no signs for e-Optha dataset. The total
the filter of order n with distance D(u,v) and a cut-off
number patch is divided into 60% training and 40% testing for
frequency at a distance D0 is given by:
validation are shown in Table II and Table III:

1
𝐻(𝑢, 𝑣) = (1)
1+[𝐷0 /D(u,v)]2𝑛

222
2019 9th IEEE International Conference on Control System, Computing and Engineering (ICCSCE), 29 Nov.–1 Dec. 2019, Penang, Malaysia

TABLE II. STATISTIC INFORMATION OF SIGN PATCHES OF DIARETDB1 Solver Method SGDM
Type of Microaneurysm Haemorrhage Exudate No- Filter Size Step Down (11,8,7,6,5,4,3,2,1)
Dataset Sign Feature Map Step Up (64,128,256,512)
Training 3341 1997 2890 3986 Activate Function ReLU, Softmax
Testing 2228 1331 1926 2656
Total Number 5569 3328 4816 6642 IV. PERFORMANCE EVALUATION
TABLE III. STATISTIC INFORMATION OF SIGN PATCHES OF E-OPHTA After In medical, the performance of the system will be
Type of Microaneurysm Haemorrhage Exudate No- evaluated based on accuracy. Accuracy of a clinical test refers
Dataset Sign to how correct a diagnostic test identifies and excludes a given
Training 1375 N/A 2516 4118 condition. The performance validation is given as follows:
Testing 917 N/A 1677 2745 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃+𝑇𝑁
(2)
Total Number 2292 N/A 4193 6863 𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁

where TP, TN, FP and FN are true positive, true negative, false
D. CNN - Patch Training positive and false negative respectively.
The CNN will be implemented on MATLAB 2017b
platform supported by NVIDIA's GTX980 graphics V. RESULTS AND DISCUSSION
processing unit (GPU). To design CNN network architecture
in Fig. 4, four (4) convolution layer were designed using 64 The fundus image has been successfully enhanced by first
feature maps in each convolution layer by kernel size of (3x3) applying the high-pass filter followed by histogram
pixels. The rectified liner unit (ReLU) activate function was equalization method. This resulted image gives clearer and
used in this research to avoid saturation and allows it to more visible pathological signs compared to the original
compute faster than sigmoid. Max-Pooling reduce the size of image. This can be observed in Fig. 5.
feature map with kernel (2x2) pixel to simplify the output,
reducing number of parameters that network need to learn.
The output of Max-Pooling was normalized by normalization
layer (NM) to speed up network training and reduce the
sensitivity to network initialization. Last 512 feature were
extracted from last Max-Pooling Layer connected together
using Fully-Connected (FC) layer layer become 512 neuron
link together. The softmax activated function normalized
output result of FC will determine which classification
probabilities by the classification layer. Fig. 5. Original image and the result after applying high-pass filter and
histogram equalization.

For this experiment 60% of total patch are used for

training, while the remaining patches are used for testing. Fig.
6 shows the training accuracy and error loss of the network
against epoch for the enhancement images.

Fig. 4. Convolution neural network architecture design. CONV: convolution,

MP: max pooling, FC: fully connected layer.

To specify training option the CNN in MATLAB, default

parameter was setup shown in Table IV:

TABLE IV. CNN SETUP DETAILS

CNN parameter Value Fig. 6. Training accuracy and error loss of the network against epoch for
the enhancement images.
Initial Learning Rate 0.0001
Maximum Epochs 1000 It can be observed that the training accuracy percentage
Validation Frequency 100 iterations and error loss have improved dramatically at early stage until
Mini Batch Size 32 the 9th epoch. Then, the accuracy percentage and error loss
Convolution layer 4 were slowly converged to more than 90% and less than 0.3
respectively. This prove that the CNN is able to extract and

223
2019 9th IEEE International Conference on Control System, Computing and Engineering (ICCSCE), 29 Nov.–1 Dec. 2019, Penang, Malaysia

learn important features required to classify the pathological image enhancement technique helps in improving visual
signs from the given dataset. quality of the fundus image as well as highlighting the
The CNN performance are further measured using pathological signs. The CNN architecture implemented in
testing dataset. The detailed test performance of the CNN is this study has achieved good classification accuracy for both
presented in confusion matrix. The confusion matrix gives DIARETDB1 and e-Ophtha datasets. The classification
detail information about the classification accuracy for each accuracy for DIARETDB1 and e-Ophtha has improved by
classes after the training process. Fig.7 shows an example of 3.6% and 2.5% respectively as compared to those images
confusion matrix for DIARETDB1 with image enhancement. without enhancement. Future work will consider other
enhancement method to be investigated for improve overall
accuracy when tested with multiple pathological signs.
ACKNOWLEDGMENT
Author would like to acknowledge Faculty of Electrical
Engineering, UiTM Shah Alam and Penang Branch,
Department of Ophthalmology, and Faculty of Medicine,
UiTM Sungai Buloh for supporting this research.

REFERENCES

[1] D. M. Squirrell and J. F. Talbot, “Screening of Diabetic Retinopathy,”

J. R. Soc. Med., vol. 96, no. 6, pp. 273–276, 2011.
[2] R. S. Biyani and B. M. Patre, “Algorithms for red lesion detection in
Diabetic Retinopathy: A review,” Biomed. Pharmacother., vol. 107,
no. July, pp. 681–688, 2018.
[3] Ministry of Health Malaysia, “Diabetic retinopathy screening: Training
module for healthcare providers,” p. 88, 2017.
Fig. 7. Confusion Matrix for DIARETDB1 with image enhancement. [4] J. I. Orlando, E. Prokofyeva, M. del Fresno, and M. B. Blaschko, “An
ensemble deep learning based approach for red lesion detection in
Table V shows overall CNN classification performance fundus images,” Comput. Methods Programs Biomed., vol. 153, pp.
115–127, 2018.
for E-Optha and DIARETDB1 with and without [5] G. G. Gardner, D. Keating, T. H. Williamson, and A. T. Elliott,
enhancement. It can be observed that image enchancement “Automatic detection of diabetic retinopathy using an artificial neural
has significantly improved the overall classification accuracy network: A screening tool,” Br. J. Ophthalmol., vol. 80, no. 11, pp.
percentage by 2.5% and 3.6% for E-Optha and DIARETDB1 940–944, 1996.
[6] K. Noronha, U. R. Acharya, K. P. Nayak, S. Kamath, and S. V.
respectively. The image enhancement improved the image Bhandary, “Decision support system for diabetic retinopathy using
quality by enchancing visual appearance of the pathological discrete wavelet transform,” Proc. Inst. Mech. Eng. Part H J. Eng.
signs. This helps the CNN to easily extract and learn the Med., vol. 227, no. 3, pp. 251–261, 2013.
important features which lead to better classification [7] B. Dupas et al., “Evaluation of automated fundus photograph analysis
algorithms for detecting microaneurysms, haemorrhages and exudates,
accuracy. Better accuracy percentage was achieved for and of a computer-assisted diagnostic system for grading diabetic
exudate and no-sign among all classes. Exudates appear in retinopathy,” Diabetes Metab., vol. 36, no. 3, pp. 213–220, 2010.
bright white or yellowish-white colour and no-sign in bright [8] P. Khojasteh, B. Aliahmad, and D. K. Kumar, “Fundus images analysis
grey which significantly differ from other pathological signs. using deep features for detection of exudates, hemorrhages and
microaneurysms,” BMC Ophthalmol., vol. 18, no. 1, pp. 1–13, 2018.
Meanwhile the lower accuracy percentage was observed for [9] P. Chudzik, S. Majumdar, F. Calivá, B. Al-Diri, and A. Hunter,
HEM and MA. This is due to the similarity in morphological “Microaneurysm detection using fully convolutional neural networks,”
shape and intensity for both signs. CNN achieved higher Comput. Methods Programs Biomed., vol. 158, pp. 185–192, 2018.
accuracy for E-Optha dataset since it has less class compared [10] L. Dai et al., “Clinical report guided retinal microaneurysm detection
with multi-sieving deep learning,” IEEE Trans. Med. Imaging, vol. 37,
to DIARETDB1. no. 5, pp. 1149–1161, 2018.
[11] S. S. Rahim, C. Jayne, V. Palade, and J. Shuttleworth, “Automatic
TABLE V. CNN CLASSIFICATION PERFORMANCE FOR E- detection of microaneurysms in colour fundus images for diabetic
OPTHA AND DIARETDB1 WITH AND WITHOUT IMAGE retinopathy screening,” Neural Comput. Appl., vol. 27, no. 5, pp. 1149–
ENCHANCEMENT
1164, 2016.
Dataset
Image Pre- Accuracy (%) [12] P. N. S. Kumar, R. U. Deepak, A. Sathar, V. Sahasranamam, and R. R.
Processing Exudate HEM MA No Sign Overall
Kumar, “Automated Detection System for Diabetic Retinopathy Using
Enhancement 99.3 N/A 99.1 99.8 99.8 Two Field Fundus Photography,” Procedia Comput. Sci., vol. 93, no.
E-Optha September, pp. 486–494, 2016.
Without
Enhancement
98.9 N/A 89.8 98.6 97.3 [13] M. R. K. Mookiah, U. R. Acharya, C. K. Chua, C. M. Lim, E. Y. K.
Ng, and A. Laude, “Computer-aided diagnosis of diabetic retinopathy:
Enhancement 96.1 93.4 95.3 99.2 96.4 A review,” Comput. Biol. Med., vol. 43, no. 12, pp. 2136–2155, 2013.
DIARETDB1
Without
[14] R. F. Mansour, “Evolutionary Computing Enriched Computer-Aided
Enhancement
97.4 89.4 84.9 95.4 92.8 Diagnosis System for Diabetic Retinopathy: A Survey,” IEEE Rev.
Biomed. Eng., vol. 10, pp. 334–349, 2017.
[15] T. Kauppi et al., “DIARETDB1-standard diabetic retino-pathy
VI. CONCLUSION database,” IMAGERET—Optimal Detect. Decis. Diagnosis Diabet.
This paper has successfully developed an automatic and Retin., pp. 15.1-15.10, 2007.
intelligent machine learning algorithm for the detection of [16] E. Decencière et al., “TeleOphta: Machine learning and image
processing methods for teleophthalmology,” Irbm, vol. 34, no. 2, pp.
Diabetic Retinopathy (DR) in fundus image. The proposed 196–203, 2013.

224
2019 9th IEEE International Conference on Control System, Computing and Engineering (ICCSCE), 29 Nov.–1 Dec. 2019, Penang, Malaysia

[17] A. Dogra and P. Bhalla, “Image sharpening by Gaussian and

Butterworth high pass filter,” Biomed. Pharmacol. J., vol. 7, no. 2, pp.
707–713, 2014.
[18] M. Ahmad, N. Zaman, and M. Al-Amin, “An experimental research in
health informatics for enhancing ovarian cancer identification in
ovarian imaging analysis using fuzzy histogram equalization,” J. Med.
Imaging Heal. Informatics, vol. 7, no. 6, pp. 1385–1390, 2017.
[19] R. Abbasi, Z. Wang, and B. Luo, “Dynamic weighted histogram
equalization for contrast enhancement using for Cancer Progression
Detection in medical imaging,” no. 1, pp. 3–8.
[20] W.-Y. Hsu and C.-Y. Chou, “Medical image enhancement using
modified color histogram equalization,” J. Med. Biol. Eng., vol. 35, no.
5, pp. 580–584, 2015.

225

View publication stats