Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

e-ISSN: 2582-5208

International Research Journal of Modernization in Engineering Technology and Science


( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:04/April-2023 Impact Factor- 7.868 www.irjmets.com

OCULAR DISEASE IDENTIFICATION USING DEEP LEARNING


TECHNIQUES
Palla Reshma*1, Sakalabhaktula Saitejesh*2, Voona Shashank*3,
Rella Jhansi*4, V.Ashok Kumar*5
*1,2,3,4Student, Department Of Electronics And Communication Engineering, Aditya Institute Of
Technology And Management, Tekkali, Srikakulam-AP, India.
*5Professor, Department Of Electronics And Communication Engineering, Aditya Institute Of
Technology And Management, Tekkali, Srikakulam-AP, India.
DOI : https://www.doi.org/10.56726/IRJMETS35505
ABSTRACT
Ocular diseases such as Glaucoma, Diabetic Retinopathy, and Cataract are among the most prevalent eye
diseases in the world. In this study, we propose a new method for classifying these diseases using fundus
images. Our method utilizes Convolutional Neural Networks (CNN), Transfer Learning (TL), and fine-tuning
techniques to improve accuracy. We collected a dataset of 4000 fundus images from publicly available sources
including Ocular Recognition, HRF, DRIVE, and IDRID, and validated our method on 400 images, with 100
images for each disease class. Our results show that our proposed method achieved an accuracy of 93.8%. This
study presents a promising approach to accurately classify ocular diseases using fundus images, which could
aid in early detection and treatment.
Keywords: Cataract, Diabetic Retinopathy, Glaucoma, Transfer Learning, Ocular Recognition, Convolution
Neural Networks.
I. INTRODUCTION
According to the first report of the World Health Organization in 2021, it is estimated that there are one
thousand two hundred million persons with an eye condition.Mostly identified eye diseases are Cataract,
Glaucoma, Diabetic Retinopathy.
fundus eye image:
The fundus is the inside, back face of the eye. It's made up of the retina and blood vessels. With fundus
photography, a special fundus camera points through the pupil to the reverse of the eye and takes filmland.

Fig.1. Retina fundus eye image


Cataract
A cataract is formation of cloudy layer on lens and causes a blurry vision.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[742]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:04/April-2023 Impact Factor- 7.868 www.irjmets.com

Fig.2. Difference between Normal and Cataract Eye


Diabetic Retinopathy
Diabetic retinopathy is caused due to damage of blood vessels in retina as shown in fig 3.

Fig.3.Difference between Normal and Diabetic Retinopathy


Glaucoma
A type of retinal condition called glaucoma causes damage to the optic nerve, which impairs vision. Most
frequently, this occurs as a result of an increase in intraocular tension. Aqueous humour, which is produced by
the eye, is secreted into the posterior chamber of the eye by the ciliary body. It enters the anterior chamber
through the pupil situated between the cornea and iris. There is a balance between production and outflow in a
healthy eye. In glaucoma suffers, the outflow filter is slightly or completely occluded. The fluid accumulates in
the chambers, increasing the strain on the heart.

Fig.4. Difference between Normal eye and Glaucoma eye.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[743]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:04/April-2023 Impact Factor- 7.868 www.irjmets.com

Fig.5. Difference between Normal vision and vision of a person suffering withan ocular diseases.
II. METHODOLOGY
The methodology for detecting Glaucoma, Cataract and Diabetic Retinopathy from fundus images is mainly
divided into 4 steps:
(1) Dataset Preparation
Identify the types of eye diseases you want to recognize: The first step is to decide which eye diseases you want
your model to identify. Some examples might include glaucoma, cataracts, macular degeneration, or diabetic
retinopathy.
Collect images of eyes: The next step is to collect a large dataset of images of eyes with and without the diseases
you want to recognize. You can use public datasets like EyePACS, DRIMDB, or Kaggle’s Diabetic Retinopathy
Detection dataset.
Table 1. Description of the dataset.
Class Name Training Images Testing Images Validation Images Total
Cataract 700 150 150 1000
Diabetic
700 150 150 1000
Retinopathy
Glaucoma 700 150 150 1000
Normal 700 150 150 1000

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[744]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:04/April-2023 Impact Factor- 7.868 www.irjmets.com

Fig. 6. Samples of Cataract, Diabetic Retinopathy, Glaucoma, and Normal


(2) Building the model
In deep learning, a convolutional neural network (CNN) is a class of deep neural networks most commonly
applied to analysing visual imagery [8]. The input layer takes 250x250 RGB images. The first 2D convolution
layer shifts over the input image using a window of size 5x5 pixels to extract features and save them on a multi-
dimensional array; in my example, the number of filters for the first layer equals 32, so the cube is (250, 250,
32). After each convolution layer, a rectified linear activation function (ReLU) is applied.

Activation has the authority to decide if a neuron needs to be activated or not by measuring the weighted sum.
ReLU returns the value provided as input directly, or the value 0.0 if the input is 0.0 or less. Because rectified
linear units are nearly linear, they preserve many of the properties that make linear models easy to optimize
with gradient-based methods. They also preserve many of the properties that make the linear model generalize
well. Transfer learning, fine tuning techniques, and ResNet, Efficient Net, and VGG deep learning pretrained
models involved in our project; Soft Max activation layer used in the output.


(1)

The model has a total of 11,184,179 parameters out of which 11,093,804 are trainable and 90,375 are non-
trainable parameters. Finally, in compilation step, the model is compiled using Adam optimizer whose learning
rate is set to 0.001 and loss is set to categorical cross entropy.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[745]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:04/April-2023 Impact Factor- 7.868 www.irjmets.com

Fig.7. Proposed Model Architecture


(3) Training the model
Load the preprocessed dataset you prepared for eye disease recognition. Divide the data into training,
validation, and testing sets.
Compile the model by specifying the loss function, optimizer, and evaluation metric. The loss function
measures how well the model is able to predict the correct class, the optimizer updates the model weights
during training, and the evaluation metric measures the performance of the model during training and
validation.
Train the model using the training set. During training, the weights of the pre-trained network are updated
based on the new data. Use a suitable batch size, number of epochs, and validation split.
Evaluate the model on the validation set. This will help you identify any issues with overfitting or under fitting.
Adjust the hyper parameters and fine-tune the model if necessary.
Test the model on the testing set. Use metrics like accuracy, precision, recall, and F1-score to evaluate the
performance of your model.
(4) Prediction
Load the trained model you saved during training, Load the image you want to make a prediction on. The image
should be preprocessed in the same way as the training images, including resizing, normalization, and any
other necessary preprocessing.
Use the predict() method of the loaded model to make a prediction on the image. The output will be a
probability distribution over the classes of eye diseases the model was trained to recognize.
Interpret the prediction by selecting the class with the highest probability. This is the predicted class for the
image. You can also calculate other metrics such as accuracy, precision, recall, and F1-score to evaluate the
performance of the model. Visualize the prediction by displaying the image and the predicted class. You can
also display the probability distribution over the classes.
III. MODELING AND ANALYSIS
The CNN architecture used for training the binary classifier is finalized with following layers after trying with
different combinations.
Input Layer: The input layer in CNN is given image data. The image data is in the form of a 3-Dimensional pixel
matrix. Input Image: The first step in the architecture of a CNN is to input an image. This image can be of any
size, but it is often preprocessed to a standard size to ensure consistency in the input data.
www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science
[746]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:04/April-2023 Impact Factor- 7.868 www.irjmets.com
Pretrained Model: The next step in the architecture of a CNN is to use a pretrained model. This model is
typically trained on a large dataset of images to learn how to recognize different features in images. This step
saves a lot of time and computational resources compared to training the model from scratch.
Activation of ReLU: After the image is input and the pretrained model is loaded, the activation function ReLU
(Rectified Linear Unit) is applied to the output of the convolutional layers. ReLU is a common activation
function used in deep learning models that helps to introduce non-linearity into the model, allowing it to better
fit complex data.
Activation of Softmax: Finally, the output of the ReLU activation is passed through a softmax function, which
normalizes the output and converts it into a probability distribution over the different possible classes that the
image may belong to. This step helps to provide a more interpretable output, allowing us to understand the
degree of confidence the model has in its classification decision.

Fig.8.Flow Chart
Accuracy alone is insufficient to evaluate the model's efficiancy. Our model and three pretrained deep learning
models are evaluated using a variety of measures, including Accuracy, Recall, Precision, and F1-Score., as
accuracy(acc) = , precision(P) = , recall(R)= , F1-Score = . True-positive, true-
negative, false-positive, and false-negative samples, respectively are indicated by the letters TP, TN, FP, and FN.
Accuracy and F1-Score are successful measures produced by averaging the performance verification findings.
The correctness indicates how exact the numbers are anticipated to be. Precision determines how well a
measurement can be reproduced or anticipated. Recall determines the correct outcome. Because there is an
exchange between accuracy and recall, a corresponding harmonic mean, known as the F1-score, is employed.
IV. RESULTS AND DISCUSSION
The performance of the proposed EfficientNetB3 is discussed in this section. Using the same dataset, our model
and two additional pre-trained algorithms were tested for cataract, diabetic retinopathy, and glaucoma
diagnosis. The experimental Results are compared to cutting-edge approaches cataracts, diabetic retinopathy,
and glaucoma.
Table 2. Comparison between the performance of the proposed three pre-trained models.
Methods Accuracy (%) Precision (%) Recall (%) F1-Score (%)
VGG19 87.6 87.6 87.5 87.5
ResNet50 85.4 86.1 85.3 85.4
EfficientNetB3 87.3 87.7 87.2 87.4
EfficientNetB3
92.4 92.3 92.3 92.3
with Fine Tuning
www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science
[747]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:04/April-2023 Impact Factor- 7.868 www.irjmets.com
A. PERFORMANCE OF EfficientNetB3
On the same dataset, we create two pre-trained CNN models (with high classification results) to assess how
well EfficientNetB3 performs. These models are (i) VGG-19 [29], and (ii) ResNet50[30]. In Table 2, the
effectiveness of our EfficientNetB3 is evaluated in terms of Accuracy, Precision, Recall, and F1-Score in
comparison to various pre-trained models. The best performance is highlighted in bold. In all evaluation
metrics, our suggested EfficientNetB3 takes first place and performs at the highest level. In comparison to the
remaining pre-trained models, EfficientNetB3 exceeds them with an accuracy of 87.3%.
Fig. 6 shows the EfficientNetB3's confusion matrix. A predicted label and an actual label, respectively, are
assigned to each row and column of the table. Out of 633 valid images 48 are wrongly classified. From fig. 6 we
can see 10 cataract images are wrongly classified as Glaucoma and normal, 2 Diabetic Retinopathy are wrongly
classified as normal, 21 Glaucoma images are wrongly classified as Cataract and normal, and 15 normal images
are wrongly classified as Cataract, Diabetic Retinopathy and normal.
In fig. 7, illustrate the validation loss, accuracy, precision and recall curves of the EfficientNetB3 model
respectively. The number of epochs is represented on the X-axis, and the loss, accuracy, precision, and recall
values are on the Y-axis. These graphics demonstrate the validation and training loss curves have merge
together after 6th epoch whereas, accuracy, precision and recall of validation and training curves become
parallel after 6th epoch.
B. COMPARISON
Table 3 summarises the reported studies on deep learning-based cataract, diabetic retinopathy, and glaucoma
detection and provides the studies' accuracy, precision, recall, and f1-score. High values for recall, accuracy,
precision, and f1-score have been bolded. Our model has the high accuracy, precision, recall, and f1-score
compared to the previous works. Our model is performing poorly on glaucoma and has less accuracy when
compared to the. However, when compared to other models, this model outperforms them in terms of accuracy,
recall, and f1-score. By Deep learning-based classifiers to detect cataract with accuracy of 92%. Deep learning-
based algorithms were used to identify detect diabetic retinopathy with about 85% accuracy. By using Machine
learning-based classifiers to detect Glaucoma with accuracy of 81%. Table 4 summarizes the predictions of the
proposed model on a new dataset.

Fig.9. The proposed model Confusion Matrix

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[748]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:04/April-2023 Impact Factor- 7.868 www.irjmets.com

Fig.10. Loss, Accuracy, Precision and Recall


Table.3. Comparison between our model and other approaches in the available literature
Author Ocular Disease Accuracy (%) Precision (%) Recall (%) F1-Score (%)
[19] Cataract 92 90 89.6 90
Diabetic
[20] 85 67 55.6 59.6
Retinopathy
[21] Glaucoma 81 73 65 69
Cataract, 96.2 89.2 96.1 92.6

Diabetic 99.3 100 99.4 99.7


This Work
Retinopathy,

Glaucoma 76 91.2 75.5 82.6


This Work Average 92.4 92.3 92.3 92.3
Table. 4. Predictions on new data
Approaches Dataset Accuracy F1-score
This work Kaggle [32] 90.75 90.78
V. CONCLUSION
In this study, we introduced Efficient net B3, a deep learning-based automated system for the identification of
glaucoma, diabetic retinopathy, and cataracts. The dataset used to feed the deep network was initially
improved by pre-processing, rearranging, and augmenting a dataset of fundus images. The algorithm Efficient
Net B3 was designed with the primary purpose of investigating various layers, activation functions, loss
functions, and optimization techniques for decreasing computing cost while maintaining model validity. In
comparison, the Efficient net B3 beat two of the three pre-trained CNN models, VGG19 and ResNet50. In terms
of accuracy (91.50%), precision (92.69%), recall (91.50%), and f1-score (91.56%), our model beat state-of-the-
art cataract, diabetic retinopathy, and glaucoma detection techniques. The ophthalmologists could quickly and
more accurately diagnose cataract, diabetic retinopathy, and glaucoma disease utilising Efficient net B3 because
to its excellent accuracy, cost- and time-efficiency. The disease like myopia, AMD, etc. cannot be detected by our
model, but it can detect cataract, glaucoma, and diabetic retinopathy. The severity of the sickness may also be
implemented in the future.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[749]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:04/April-2023 Impact Factor- 7.868 www.irjmets.com
VI. REFERENCES
[1] R.R. Bourne,S.R. Flaxman,T. Braithwaite,M.V. Cicinelli,A. Das,J.B. Jonas,J. Keeffe,J.H. Kempen,J. Leasher,H.
Limburg etal.," Magnitude, temporal trends, and protrusions of the global frequence of blindness and
distance and near vision impairment a methodical review and meta- analysis," The Lancet Global
Health,vol. 5,no. 9, pp. e888 – e897, 2017.
[2] Güven,A. Automatic discovery of age- related macular degeneration pathologies in retinal fundus
images. Comput. styles Biomech. Biomed.Eng. 16( 4), 425 – 434( 2013)
[3] B. Raju, N. S. D. Raju, J. Akkara, and A. Pathengay, ‘‘Do it yourself smartphone fundus camera—
DIYretCAM,’’ Indian J. Ophthalmol., vol. 64, no. 9, p. 663, 2016.
[4] Dong,Y., Wang,Q., Zhang,Q., Yang,J. Bracket of cataract fundus image grounded on retinal vascular
information. In International Conference on Smart Health, Springer,pp. 166 – 173( 2016)
[5] Addict,W., Shen,R., Zhang,Q., Yang,J.J., Li,J. star element analysis- grounded cataract grading and
bracket. In 2015 17th International Conference onE-health Networking, operation and Services(
HealthCom), IEEE,pp. 459 – 462( 2015)
[6] Manchalwar,M., Warhade,K. Discovery of cataract and conjunctivitis complaint using histogram of
acquainted grade. Int.J.Eng. Technol.( IJET)( 2017)
[7] Qiao,Z., Zhang,Q., Dong,Y., Yang,J.J. Application of SVM grounded on inheritable algorithm in bracket of
cataract fundus images. In 2017 IEEE International Conference on Imaging Systems and ways( IST),
IEEE,pp. 1 – 5( 2017)
[8] Xiong,L., Li,H., Xu, L. An approach to estimate blurriness in retinal images with vitreous nebulosity for
cataract opinion.J. Healthc.Eng. 2017( 2017)
[9] Krizhevsky,A., Sutskever,I., Hinton,G.E. Imagenet bracket with deep convolutional neural networks. In
Advances in Neural Information Processing Systems,pp. 1097 – 1105( 2012 10) Pan,S.J., Yang, Q. A
check on transfer literacy. IEEE Trans. Knowl. DataEng. 22
[10] Pan, S.J., Yang, Q.: A survey on transfer learning. IEEE Trans. Knowl. Data Eng. 22(10), 1345–1359
(2009)
[11] A. Budai, R. Bock, A. Maier, J. Hornegger, and G. Michelson, ‘‘Robust vessel segmentation in fundus
images,’’ Int. J. Biomed. Imag., vol. 2013, pp. 1–11, Dec. 2013.
[12] C. Hernandez-Matas, X. Zabulis, A. Triantafyllou, P. Anyfanti, S. Douma, and A. A. Argyros, ‘‘FIRE: Fundus
image registration dataset,’’ Model. Artif. Intell. Ophthalmol., vol. 1, no. 4, pp. 16–28, 2017.
[13] Z. Zhang, F. S. Yin, J. Liu, W. K. Wong, N. M. Tan, B. H. Lee, J. Cheng, and T. Y. Wong, ‘‘ORIGA-light: An
online retinal fundus image database for glaucoma analysis and research,’’ in Proc. Annu. Int. Conf.
IEEE Eng. Med. Biol., Aug. 2010, pp. 3065–3068.
[14] P. Porwal, S. Pachade, R. Kamble, M. Kokare, G. Deshmukh, V. Sahasrabuddhe, and F. Meriaudeau,
‘‘Indian diabetic retinopathy image dataset (IDRiD): A database for diabetic retinopathy screening
research,’’ Data, vol. 3, no. 3, p. 25, Sep. 2018.
[15] E. Decencière, G. Cazuguel, X. Zhang, G. Thibault, J.-C. Klein, F. Meyer, B. Marcotegui, G. Quellec, M.
Lamard, R. Danno, D. Elie, P. Massin, Z. Viktor, A. Erginay, B. Laÿ, and A. Chabouis, ‘‘TeleOphta: Machine
learning and image processing methods for teleophthalmology,’’ IRBM, vol. 34, no. 2, pp. 196–203, Apr.
2013.
[16] J. Staal, M. D. Abr{moff, M. Niemeijer, M. A. Viergever, and B. van Ginneken, ‘‘Ridge-based vessel
segmentation in color images of the retina,’’ IEEE Trans. Med. Imag., vol. 23, no. 4, pp. 501–509, Apr.
2004.
[17] A. A. Jeny, A. N. M. Sakib, M. S. Junayed, K. A. Lima, I. Ahmed, and M. B. Islam, ‘‘SkNet: A convolutional
neural networks-based classification approach for skin cancer classes,’’ in Proc. 23rd Int. Conf. Comput.
Inf. Technol. (ICCIT), Dec. 2020, pp. 1–6.
[18] Richard Bina Jadi Simanjuntak, Yunendah Fu’adah, Rita Magdalena, Sofia Saidah, “Cataract
Classification Based on Fundus Images Using Convolutional Neural Network,” JOIV, pp.33-38, Mar.
2022.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[750]
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:04/April-2023 Impact Factor- 7.868 www.irjmets.com
[19] Zubiar Khan, Fiaz Gul Khan, Ahmad Khan, Zia Ur Rehman, “Diabetic Retinopathy Detection Using VGG-
NIN a Deep Learning Architecture,” IEEE, vol.9, pp.61408-61416, Apr. 2021.
[20] Danilo Leite, Maria Campelos, Ana Fernandes, Pedro Batista, “Machine Learning automatic assessment
for glaucoma and myopia based on Corvis ST data,” ELSEVIER, pp.454-460, 2021.
[21] J. Y. Choi, T. K. Yoo, J. G. Seo, J. Kwak, T. T. Um, and T. H.Rim, “Multi-categorical deep learning neural
network to classify retinal images: A pilot study employing small database,” PloS one, vol. 12,no. 11, p.
e0187336, 2017.
[22] R. Arunkumar and P. Karthigaikumar, “Multi-retinal disease classification by reduced deep learning
features,” Neural Computing and Applications, vol. 28, no. 2, pp. 329–334, 2017.
[23] R. Arunkumar andP. Karthigaikumar, “Multi-retinal complaint bracket by reduced deep literacy
features, ” Neural Computing and Applications,vol. 28,no. 2,pp. 329 – 334, 2017.
[24] V. Gulshan,L. Peng,M. Coram,M.C. Stumpe,D. Wu,A. Narayanaswamy,S. Venugopalan,K. Widner,T.
Helpmates,J. Cuadros etal., “ Development and confirmation of a deep literacy algorithm for discovery
of diabetic retinopathy in retinal fundus photos, ” Jama,vol. 316,no. 22,pp. 2402 – 2410, 2016. 24)Z.
Li,Y. He,S. Keel,W. Meng,R.T. Chang, andM. He, “ efficacity of a deep literacy system for detecting
glaucomatous optical neuropathy grounded on color fundus photos, ” Ophthalmology,vol. 125,no. 8,pp.
1199 – 1206, 2018.
[25] J.J. G´omez- Valverde,A. Ant´on,G. Fatti,B. Liefers,A. Herranz,A. Santos,C.I. S´anchez, andM.J. Ledesma-
Carbayo, “ Automatic glaucoma bracket using color fundus images grounded on convolutional neural
networks and transfer literacy, ” Biomedical optics express, vol. 10,no. 2,pp. 892 – 913, 2019.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science


[751]

You might also like