Professional Documents
Culture Documents
Diabetic - Retinopathy - Muruga Radha Devi D
Diabetic - Retinopathy - Muruga Radha Devi D
Department of CSE,
Department of Information Technology
Karunya Institute of Technology and Sciences
Sri SaiRam Institute of Technology,
Coimbatore.
Chennai, Tamil Nadu ,India. iwinjoseph@karunya.edu
leela.it@sairamit.edu.in
5. C. Prasanna Lakshmi
Department of Information Technology
Sri SaiRam Institute of Technology,
Chennai, Tamil Nadu ,India.
Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 06,2022 at 06:47:10 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Second International Conference on Smart Electronics and Communication (ICOSEC).
IEEE Xplore Part Number: CFP21V90-ART; ISBN: 978-1-6654-3368-6
Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 06,2022 at 06:47:10 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Second International Conference on Smart Electronics and Communication (ICOSEC).
IEEE Xplore Part Number: CFP21V90-ART; ISBN: 978-1-6654-3368-6
Fig.2: Architecture of the automatic DR stages classification Model Fig.4: Preprocessing of the OCT Retinal Image
OCT Dataset
The OCT ret inal Data set is splitted into Test set Feature Extraction and Classification using CNN Model
(20%) and remain ing 80% of OCT retinal images are further The convolutional Neural network is imp lemented using
divided into validation set (20%) and Training set (80%). The keras API. Sequential model is used for this classification.
set comprising training data and validation data are used for This model uses six alternate layers of convolution layers and
training the CNN Model and the test set is used for calculat ing Pooling layers are used. Initial Convolution layer co mprises
the performance of the CNN model. The Splitting of the filters to mine basic attributes of the images. The remain ing
Dataset set is shown in figure 3. layers extract co mplex features that are needed for the
classification of DR. Pooling layers perform down sampling
which retains the information while reducing the special
OCT resolution. In our model, Max-pooling is used. The activation
Retinal function (ReLu ) is used in our model. To avoid overfitting,
Data set drop out layer is added. Flattening layers are added to convert
2D into 1D vectors. This model uses 3 fully connected layers
80% Training (dense). The first two dense layer uses ReLu activation
Set 20% Test Set function and the last dense layer includes soft-max act ivation
function. Two stages namely classificat ion stage-1 and
classification stage-2 are used in our model. Classification
80% Training 20% Validation stage-1 classifies whether the ret inal image is either with DR
Set Set or not. The classification stage-2 classifies the DR image into
four grades namely NPDR-mild, moderate, severe and PDR.
Fig.3. Splitting of OCT Retinal Data Set Model summary of classification stage-1 is shown in Figure 5.
Preprocessing
Input OCT images are segmented into 12 layers using
Markov Gibbs Random Fields (M GRF) method. The Fovea is
located in the center of the image. All the 12 layers are not of
same thickness. Nearly 5 layers are merged in the fovea
region. Fovea region is also very much needed for the
detection of DR. Three patches namely (temporal patches,
Nasal patches and foveal patches) are ext racted as in figure 4.
Three patches contain information related to reflectivity,
thickness and curvature. The thickness of each retinal layers
detects the type of retinal diseases. Reflectivity of damaged
retinal varies fro m normal ret ina. These features are included
for identify ing the occurrence of microaneurysm, hard
exudates, hemorrhage in OCT images. The performance of
CNN model improves when the training data s et increases.
Data augmentation is used to increase the training images.
Operations like flipping, shifting and rotation are used for data
augmentation. These operations are included on the fly during
processing hence it does not occupy more space. In OCT ,the
Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 06,2022 at 06:47:10 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Second International Conference on Smart Electronics and Communication (ICOSEC).
IEEE Xplore Part Number: CFP21V90-ART; ISBN: 978-1-6654-3368-6
The train ing patches (temporal, nasal and foveal) are fed
into two classification models. The train ing accuracy of the
model is 98.97%. The classificat ion stage-1 classifies the
images with an accuracy of 96.32%. The classificat ion stage-2
classifies the stages of DR with an accuracy of 95%. The
accuracy of the classificat ion of DR grades is 95%.The
method combines the three patches namely nasal, foveal and
temporal as features in support of classification regarding DR
stages in OCT images, accuracy of the classification model
improves.The newness of the model lies in the categorizat ion
of DR in OCT retinal images using three patches which was
missing in earlier studies.
Fig.5: CNN model summary for classification Stage-1 CONCLUSION AND FUT URE SCOPE
Manual interpretation of OCT retinal images is t ime
The classification stage-2 consists of 4 alternative
consuming. More -over the interpretation of OCT retinal
convolution layers with ReLu activation function and pooling
images differ accord ing to the Ophthalmologist experience
layers (Max). Dropout layer is also included in the d irection of
and review. Th is computer aided diagnosis system that
avoiding the overfitting problem. Rectified Linear Unit
automatically classifies the OCT retinal images will help the
(ReLu ) provides the input for positive input value and zero for
patients to get the opinion. Due to the high resolution of OCT
negative input value. To stop linearity, ReLu is used. It works
imaging modality, the earlier stages of Diabetic Retinopathy
so that it will not activate the neuron equally thereby
can be accurately detected. Diabetic Retinopathy using OCT
providing a Non-Linear output at each level. Three fu lly
imaging modality need further exploration in the area of
connected layers are used for classificat ion. Activation
research so as to classify the earlier stages of DR in precise.
function Sig moid is used in the final dense layer which
performs mult iclass classification. OCT ret inal images are
classified into four levels as mild, moderate, severe and REFERENCES
PDR.Learning algorith m used in our model is Adam and the
1. Sohini Roychowdhury, Dara D. Koozekanani , Keshab K.
learning rate 0.001 is chosen.The loss function used is soft Parhi,”DREAM : Diabetic Retinopathy Analysis Using
max function which is very much best for mu lticlass M achine Learning ”,pp:1717 – 1728, vol.18, Iss:5,2014.
classification. 2. Yitian Zhao, Yalin Zheng, Yonghuai Liu, Jian Yang, Yifan
Zhao, Duanduan Chen, Yongtian Wang,"Intensity and
Compactness Enabled Saliency Estimation for Leakage
Detection in Diabetic and M alarial Retinopathy” “IEEE
Trans. on M ed. Imag.,vol.36, Iss:1,2017.
Performance Evaluation of Convolution Nueral Network
3. Ramon Pires, Sandra Avila, Herbert F. Jelinek, Jacques
Model Wainer , Eduardo Valle, Anderson Rocha, ”Beyond
The classification performance of CNN is calculated Lesion-Based Diabetic Retinopathy: A Direct Approach for
by considering the accuracy, sensitivity and specificity. Referral”,IEEE Journ. of Biomed. and Health Inform.,vol.
Let True.positive = PT, True.negative =NT, False.positive = 21, Iss:1,2017.
PF, False.negative =NF 4. Xianglong Zeng, Haiquan Chen,Yuan Luo,Wenbin Ye “
Then Automated Diabetic Retinopathy Detection Based on
Accuracy= (PT +NT)/(PT+NT+PF+NF) Binocular Siamese-Like Convolutional Neural
Network“,IEEE Access,Vol.7,2019.
5. Xiaomeng Li,Xiaowei Hu, Lequan Yu, Lei Zhu, Chi-Wing
Fu, Pheng-Ann Heng,” CANet: Cross-Disease Attention
Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 06,2022 at 06:47:10 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Second International Conference on Smart Electronics and Communication (ICOSEC).
IEEE Xplore Part Number: CFP21V90-ART; ISBN: 978-1-6654-3368-6
Network for Joint Diabetic Retinopathy and Diabetic Heng,”UD-M IL for OCT images Classification ” IEEE
M acular Edema Grading”, IEEE Trans. on M ed. Journ. of Bio.med. and Health
Imag.,vol.39, Iss: 5 ,2020. Inform.,vol.24, Issue: 12,2020.
6. Lifeng Qiao, Ying Zhu, Hui Zhou,” Diabetic Retinopathy 14. Qiaoliang Li, Shiyu Li, Zhuoying He,Huimin
Detection Using Prognosis of M icroaneurysm and Early Guan, Runmin Chen; Ying Xu,Tao Wang,Suwen Qi, Jun
Diagnosis System for Non-Proliferative Diabetic M ei,Wei Wang “DeepRetina: Layer Segmentation of
Retinopathy Based on Deep Learning Algorithms”, IEEE Retina in OCT Images Using Deep Learning”,
Access,2020. Translational Vision Science & Techn.,vol.9,Dec.2020.
7. Teresa Araújo, Guilherme Aresta ,Luís M endonça, Susana 15. Jefferson Alves Sousa , Anselmo Paiva , Aristofanes Silva,
Penas, Carolina M aia, Angela Carneiro ,Ana M aria Joao Dallyson Almeida , Joao Dallyson Almeida , Joao
M endonça , Aurélio Campilho ,"Data Augmentation for Otavio Diniz , Weslley Kelson Figueredo,M arcelo
Improving Proliferative Diabetic Retinopathy Detection in Gattass,”Automatic segmentation of retinal layers in OCT
Eye Fundus Images”,IEEE Access,vol.8 ,2020. images with intermediate age-related macular degeneration
8. Sehrish Qummar, Fiaz Gul Khan, Sajid Shah, Ahmad Khan using U-Net and DexiNed,2021,PLOS ONE.
,Shahaboddin Shamshirband, Zia Ur Rehman, Iftikhar 16. Karunakaran, P., and Yasir Babiker Hamdan. “Early
Ahmed Khan, Waqas Jadoon,“A Deep Learning Ensemble Prediction of Autism Spectrum Disorder by Computational
Approach for Diabetic Retinopathy Detection”, IEEE Approaches to fM RI Analysis with Early Learning
Access,vol.7,2019. Technique.” Journal of Artificial Intelligence,vol. 2, no.
9. Laxmi Gella, Rajiv Raman, Padmaja Kumari 04,pp: 207-216,2020.
Rani, and Tarun Sharma,” Spectral domain optical 17. Palani, U., M rs D. Vasanthi, and M s S. Rabiya Begam.
coherence tomography characteristics in diabetic “Enhancement of M edical Image Fusion Using Image
retinopathy”,Oman J Ophth..,2014, 7(3),pp.126–129. Processing.” Journal of Innovative Image Processing
10. M ohammed Ghazal, Samr Samir Ali, Ali H. M ahmoud, (JIIP),vol.2, no. 04 ,pp:165-174,2020.
Ahmed M . Shalaby, Ayman El-Baz, “Accurate Detection 18. Tripathi, M ilan. “Analysis of Convolutional Neural
of Non-Proliferative Diabetic Retinopathy in Optical Network based Image Classification Techniques.” Journal
Coherence Tomography Images Using Convolutional of Innovative Image Processing (JIIP),vol.3,no. 02,pp:100-
Neural Networks”,IEEE,2020. 117,2021.
11. Bashir Isa Dodo, Yongmin Li, Djibril Kaba, Xiaohui Liu 19. Dhaya, R. “Analysis of Adaptive Image Retrieval by
“Retinal Layer Segmentation in Optical Coherence Transition Kalman Filter Approach based on Intensity
Tomography Images” IEEE Access ,vol.7,2019. Parameter.” Journal of Innovative Image Processing
12. G. Roy, S. Conjeti, S. P. K. Karri, D. Sheet, A. Katouzian, (JIIP),vol.3,no.01,pp:7-20,2021.
C. Wachinger, et al., "ReLayNet/Retinal layer and fluid
segmentation of macular optical coherence tomography
using fully convolutional
networks", Biomed.Opt.Exp.,vol.8,pp. 3627-42,Aug. 2017.
13. “Xi Wang, Fangyao Tang, Hao Chen, Luyang Luo, Ziqi
Tang, An-Ran Ran, Carol Y. Cheung, Pheng-Ann
Authorized licensed use limited to: Sri Sivasubramanya Nadar College of Engineering. Downloaded on January 06,2022 at 06:47:10 UTC from IEEE Xplore. Restrictions apply.