Name: CNN Based Fundus Images Classification For Glaucoma Identification
Uploaded: 2022-01-26
Duration: 15 min 12 s
Description: CNN Based Fundus Images Classification For Glaucoma Identification.

CNN Based Fundus Images Classification For Glaucoma Identification. 

Scene 1

CNN Based Fundus Images Classification For Glaucoma Identification

Paper ID – 75 2nd International Conference on Advanced Research in Computing (ICARC 2022) 23-24, February 2022

Scene 2

Author Details: Thisara Shyamalee,  	Department of Computer Science and Engineering, University of Moratuwa, Sri Lanka Dr. Dulani  Meedeniya, 	Department of Computer Science and Engineering, University of Moratuwa, Sri Lanka

Glaucoma is on the rise!!!. Source: The World Glaucoma Association, 2020. 

Scene 3

Source: The World Glaucoma Association, 2020

million
have glaucoma
million
are predicted to have
glaucoma by 2040
every 2 - 3 years
of glaucoma is undetected in
developing countries
billi
have no access to eye-care,
due to unfair distribution
KNOW WHEN TO GET CHECKED
Because, the EARLIER the diagnosis, the less
damage will be done and the more vision will
be saved!
BEFORE AGE 40
FROM 40 - 60
AFTER 60
every 2
- 4 years
- 2 years

Introduction. Healthy eyes vision loss due to glaucoma. 

Scene 4

Healthy eyes
vision loss due to glaucoma

Glaucoma: leads to vision loss for people over 40-60 years old globally. Treatment planning relies on images,  	 such as Retinal Fundus, OCT scans  	 acquired from the optic nerve. Glaucoma can be prevented by early   	 detection and treatment.

Motivation. Computational models can be used as a second opinion in the glaucoma diagnosis process. Address the issue due to misdiagnosis Better classification with fundus images. 

Scene 5

Computational models can be used as a second opinion in the glaucoma diagnosis process. Address the issue due to misdiagnosis Better classification with fundus images

Objectives. Explore literature related to deep learning and computer vision techniques for retinal fundus images Design and develop a Convolution based fundus images classification to support glaucoma diagnosis process Analyse the accuracy of the classification results. 

Scene 6

Explore literature related to deep learning and computer vision techniques for retinal fundus images Design and develop a Convolution based fundus images classification to support glaucoma diagnosis process Analyse the accuracy of the classification results

Related Studies. Study Dataset DL model Performance Accuracy Specificity Sensitivity AUC (Diaz‑Pinto et al, 2019) Fundus VGG16 89% 88% 90% 0.96 (Syna et al, 2020) Fundus DeepLabv3+ and MobileNet 97.37% - - 1 (Diego et al., 2021) Fundus CNN 94.60% 92.50% 94.57% - ( Rehman et al., 2021) Fundus VGG16, VGG19 97.8% (Khalil et al, 2018) OCT Scans Used ML technique 90% 95% 87% -. 

Scene 7

Study Dataset DL model Performance Accuracy Specificity Sensitivity AUC (Diaz‑Pinto et al, 2019) Fundus VGG16 89% 88% 90% 0.96 (Syna   et  al, 2020) Fundus DeepLabv3+ and MobileNet 97.37% - - 1 (Diego et al., 2021) Fundus CNN 94.60% 92.50% 94.57% - ( Rehman  et al., 2021) Fundus VGG16, VGG19 97.8% (Khalil et al, 2018) OCT Scans Used  ML technique 90% 95% 87% -

Apart of the studies conducted using these dataset, novelty of our study is demonstrate new approach fine-tuning three DL models and  the high accuracy, specificity and sensitivity implied from this new approach suggest that fine -tuned CNN architectures are a robust alternative for an automatic glaucoma diagnosis.

Explore Explore existing literature, techniques and data sets as well. Preprocessing Apply the data preprocessing techniques to remove unwanted noises. Employ data augmentation techniques to reduce the chance of overfitting and to improve the generalization of the algorithm. Classification Develop a DL method to classify with different data namely fundus images into group of similar details. Analyze the normal and glaucoma eye using classifying data with best performances. Analyze. 

Scene 8

Explore Explore existing literature, techniques and data sets as well. Preprocessing Apply the data preprocessing techniques to remove  unwanted  noises. Employ data augmentation techniques to reduce the chance of overfitting and to improve the generalization of the algorithm. Classification Develop a DL method to classify with different data namely fundus images into group of similar details. Analyze the normal and glaucoma eye using classifying data with best performances. Analyze

Data Fundus Images. Pre-processing (Fundus Images) Dilation CLAHE. 

Scene 9

Pre-processing (Fundus Images) Dilation CLAHE

Deep Learning Models Inception v3 VGG19 ResNet50

Accuracy 5-fold cross validation F1-score Specificity Sensitivity AUC

Data Augmentation Rotation Shearing Zooming Horizontal flip Vertical flip Width shift range Height shift range

RIM-ONE dataset https://medimrg.webs.ull.es/ ACRIMA Dataset https://figshare.com/s/c2d31f850af14c5b5232. 

Scene 10

RIM-ONE dataset   https://medimrg.webs.ull.es/ ACRIMA Dataset https://figshare.com/s/c2d31f850af14c5b5232

Dataset Name Image type Glaucoma Normal RIM-ONE Fundus Images 399 543 ACRIMA Fundus  Images 396 309

Implementation. Modified Inception v3 architecture We utilized the Keras API's Inception v3 model, which was pre-trained on the ImageNet dataset. For the method of glaucoma diagnosis, additional layers were applied instead of the model's top dense layers. The Global Average Pooling Layer was placed after the Inception v3 model to reduce the parameters, followed by the Dense Layer (512) , and lastly added the Softmax Layer .. 

Scene 11

Modified Inception v3 architecture We utilized the Keras API's Inception v3 model, which was pre-trained on the ImageNet dataset. For the method of glaucoma diagnosis, additional layers were applied instead of the model's top dense layers.  The  Global Average Pooling Layer  was placed after the Inception v3 model to reduce the parameters, followed by the  Dense Layer (512)  , and lastly added the  Softmax Layer .

512 Dense
Global Avenge
PNIing Lay •r
Layer
Inception v3 Architecture

2. Modified VGG19 architecture In this study, pre-trained VGG19 was fine-tuned. The Global Average Pooling Layer is followed by the last three newly added layers Dropout Layer with 0.5 rate, Dense Layer (256) with ReLU activation function, and finally, Softmax Layer with two outputs.. 

Scene 12

2.	 Modified VGG19 architecture In this study, pre-trained VGG19 was fine-tuned.  The  Global Average Pooling Layer  is followed by the last three newly added layers  Dropout Layer  with  0.5  rate,  Dense Layer (256)  with  ReLU  activation function, and finally,  Softmax Layer  with two outputs.

224-224•64
224-224•3
112X112X128
S6xS6x2S6
Fine tuned
28x28xSi2
14X14X512
7X7X512
Convolution
Global Average Pooling Lover
Laver (O. S)
Softmax Laver

3.	Modified ResNet50 architecture In this, pre-trained ResNet50 was fine-tuned. The last two layers (Fully Connected and Softmax) were altered first. The Global Average Pooling Layer is followed by the Dropout Layer with 0.5 rate , Dense Layer (256) with ReLU activation function, and finally, Softmax Layer with two outputs.. 

Scene 13

3.	Modified ResNet50 architecture In this, pre-trained ResNet50 was fine-tuned. The last two layers  (Fully Connected and Softmax) were altered first.  The Global Average Pooling Layer is followed by the  Dropout Layer  with  0.5 rate ,  Dense Layer (256)  with  ReLU activation  function, and finally,  Softmax Layer  with two outputs.

7x7 Conv 64
/2
3x3
64
64
conv.
3x3 conv
64
3x3 conv
64
64
30 conv
64
conv, I
30 conv 128
3x3 conv,
128
3x3 conv 128
3x3 conv
128
3x3 co
128
3x3 Conv
128
128
conv
n conv 256 2
3x3 conv
256
3x3 conv
256
3x3 conv
256
3x3 conv
256
30 conv 256
3x3 conv 256
3x3 conv 256
30 Conv
256
30 Conv
3X3 C (HIV
256
conv 51 2
512
conv
conv, 512
512
30 Conv 512
3x3 conv 512
ave Dool
sunsau

Global Average Pooling Layer Reduce the number of parameters and the amount of computation performed in network. Dense Layer To classify the fundus image, the feature map of the final average pooling layer is converted into a one-dimensional vector and linked to one or more fully connected layers known as dense layers. Dropout Layer Drop-out layers aid in the prevention of overfitting as well as the speeding up of the training process. Softmax Layer The Softmax layer acts as an activation function, determining the probability of each class and classifying the target images.. 

Scene 14

Global Average Pooling Layer 		  Reduce the number of parameters and the amount of computation performed in 			 network. 	 Dense Layer 	 	 To classify the fundus image, the feature map of the final average pooling layer is  			converted into a one-dimensional vector and linked to one or more fully connected 		layers known as dense layers. 	 Dropout Layer 		 Drop-out layers aid in the prevention of overfitting as well as the speeding 	up of 			the training process. Softmax Layer 		 The Softmax layer acts as an activation function, determining the probability of each 		class and classifying the target images.

Results of Inception-v3 architecture on (a) RIM-ONE (b) ACRIMA. 

Scene 15

Results of Inception-v3 architecture on (a) RIM-ONE (b) ACRIMA

2 . Results of VGG19 architecture on (a) RIM-ONE (b) ACRIMA. 

Scene 16

2 .	 Results of VGG19  architecture on (a) RIM-ONE (b) ACRIMA

3.	Results of ResNet50 architecture on (a) RIM-ONE (b) ACRIMA. 

Scene 17

3.	Results of ResNet50 architecture on (a) RIM-ONE (b) ACRIMA

Summary of Results. 18. PERFORMANCE OF CNN ARCHITECTURES ON RIM-ONE CNN Architectures Performance Accuracy Precision Recall Fl score Sensitivity Specificity Loss AUC Inception-v3 96.56% 97.22% 98.31% 97.76% 98.31% 90.90% 0.0601 0.98 VGG19 94.95% 95.37% 97.98% 96.65% 97.98% 86.13% 0.0792 0.95 ResNet50 95.49% 96.50% 97.59% 97.04% 97.59 C/o 8888% 0.0689 0.97. 

Scene 18

PERFORMANCE OF CNN ARCHITECTURES ON RIM-ONE
CNN Architectures
Performance
Accuracy
Precision
Recall
Fl score
Sensitivity
Specificity
Loss
AUC
Inception-v3
96.56%
97.22%
98.31%
97.76%
98.31%
90.90%
0.0601
0.98
VGG19
94.95%
95.37%
97.98%
96.65%
97.98%
86.13%
0.0792
0.95
ResNet50
95.49%
96.50%
97.59%
97.04%
97.59 C/o
8888%
0.0689
0.97

PERFORMANCE OF CNN ARCHITECTURES ON ACRIMA
CNN Architectures
Performance
Accuracy
Precision
Recall
Fl score
Sensitivity
Specificity
Loss
AUC
Inception-v3
98.52%
98.84%
99.50%
99.17%
99.50%
90.90%
0.0414
0.99
VGG19
92.64%
93.45%
95.23%
78.78%
0.0756
0.94
ResNet50
95.58%
96.49%
98.21%
97.34%
98.21%
83.33%
0.0598
0.96

ACRIMA
RIM-ONE
Inception-v3
VGG19
ResNet50
Inception-v3
VGG19
ResNet5()
Adam
SGD
SGD
Adam
SGD
SGD
98.52
92.64
95.58
96.56
94.95
95.49

19. Comparison With The Existing S tudies. 

Scene 19

20. Future Work. Develop as Ensemble model Deploy as a tool with interactive interfaces Experiment with other possible deep learning techniques Deploy in a clinical practice as a prototype and do a user study. 

Scene 20

Develop  as Ensemble  model Deploy as a tool with interactive  interfaces Experiment  with other possible deep learning techniques Deploy  in a clinical practice as a prototype and do a user  study

Conclusion. This paper presented deep learning approaches to classify fundus images and we’ve proposed three fine-tuned architectures based on Inception-v3, VGG19, and ResNet50 to classify glaucoma and normal images. We have applied dilation and augmentation during data pre-processing. 5-fold cross-validation based results showed that Inception-v3 predicts glaucoma with an accuracy of 98.52% among the other convolutional neural networks architectures.. 

Scene 21

This paper presented deep learning approaches  to classify fundus  images and we’ve  proposed three fine-tuned architectures based on Inception-v3, VGG19, and ResNet50 to classify glaucoma and normal  images.  We  have applied dilation and  augmentation  during data pre-processing.  5-fold  cross-validation based results showed that Inception-v3 predicts glaucoma with an accuracy of 98.52% among the other convolutional neural networks architectures.

Scene 22

 RIM-ONE dataset   https://medimrg.webs.ull.es/ ACRIMA Dataset https://figshare.com/s/c2d31f850af14c5b5232 	