Materials and Method

Database and Division of Images

In this section, the deployment of the proposed method, feature extraction, and classification to diagnose glaucoma and its stages is presented. The first step is the acquisition of digital fundus images from the public databases like RIM-ONE, DRIONS-DB, DRIVE. HRF, and DRISHTI - GS. Certain images are picked up for training the CNN. For this work, a large number of datasets was collected to analyze and investigate the onset of glaucoma severity level through the DL approach. The datasets are labeled by specialists in the field and certain ground truth images were also considered. DL approach was applied on the databases to assess the viability of the technology via performance on images with varying resolutions [35]. The data acquisition resulted in И55 fundus images collected through fundus cameras, namely Canon, Nidek AFC-210, and Top-con. Images had varying resolutions, for example 1924 x 1556 pixels and 2336 x 2336 pixels. Each image class sample was classified randomly into entirely non-dependent sub-groups for the purpose of training and testing. The testing class image is invariably same for all the glaucoma grades in such way to reduce potential bias occurring due to sample size imbalance. The images were rated by experienced professionals in grading the images into different classes. Table 5.1 presents the division of acquired fundus images based on their severity level.

Preprocessing and Down-Sampling

To remove noise and unwanted backgrounds from the acquired images, the image preprocessing step is employed. Certain morphological operations like contrast enhancement, binarization, erosion, and dilation are performed to preserve the field of view (FOV) of the images [36]. Over fitting due to the small number of datasets can be avoided by data augmentation processes. The data augmentation increases the variation within a training dataset. This repetition process involves cropping, random translations, and flipping the images to a fixed downscaled size of 227 x 227 [37].

TABLE 5.1

Division of Images for Glaucoma Identification Based on Severity Level

Stages

Healthy

Glaucoma

Early

Moderate

Deep

Training

301

142

84

228

Testing

100

100

100

100

Total

401

242

184

328

Deep Learning Architecture (DLA)

A DL architectural framework is an integration of feature extraction and classification.

The strength of DLA depends on weight sharing. This is constructed with an assumption that the image may contain similar structures in various locations. Convolution filters can detect these structures with kernel functions. Non-linear activation transform, pooling function, and translation invariance are the operations that take place in the CNNs [38]. A CNN is a DL model that comprises of the following network layers: input layer, convolutional layer, batch normalization layer, pooling layer, softmax layer, etc. The size of the input decides the number of layers; not necessarily all the layers have to be used for the application. The deeper the network the better is the training, but at the same time computational time increases. One important property of CNN is that they have the ability to self-learn and self-organize unsupervised [39]. In order to obtain better training and reduce computational time, one can use the minimum number of layers to get the maximum benefit by employing efficient network parameters. By tuning the network parameters effectively it is possible to achieve a better classification rate. A brief note on the layers of CNN is given below

Input Layer

It must be present in the CNN to take the input forward to the network.2D and 3D forms of the images are accepted by the layer with dimensions being initialized.

Convolutional Layer

Convolution of the input image is performed in this layer resulting in feature maps that are used as input to the next layer.

Batch Normalization Layer

This layer ensures fast learning and allows the flow of normalized samples in between the intermediate layers, improving overall performance.

Rectified Linear Unit

This layer reduces data redundancy without losing vital information.

Max Pooling Layer

Applying the max pooling operation on each feature map reduces the size of the map according to the user-chosen value.

Fully Connected Layer

The max pooling layer donates neurons to this layer and they are connected to every neuron of this layer. The number of classification classes depends on the output of this layer.

Softmax Layer

A normalized exponential function operation called the softmax is performed which reduces the data sample dimensionality by removing the outliers so that they fall in the range 0-1. The output label is determined by this layer.

Methodology

Proposed CNN Architecture

CNNs have two parts, the convolutional layer and the fully connected layer. In the convolutional layer the input of every layer serves as the output of the preceding layer that acts as feature extractor. These extracted features are classified in the fully connected layer followed by the softmax activation function. Each class has only one output neuron. Note that by tuning the parameters misclassification error is reduced. The overview of the 25-layer CNN is shown in the Figure 5.3 and Table 5.2 presents the parameters used to construct the CNN.

Overview of proposed CNN architecture

FIGURE 5.3 Overview of proposed CNN architecture.

TABLE 5.2

Parameter Description of the Proposed CNN

S. No

Name

Type

Activations

1

Image input

227 x 227 x 3 image with 'zero center normalization’

Image Input

227 x 227 x 3

2

convl-96 11x11x3 convolution with stride [4 4] and padding [0 0 0 0]

Convolution

55 x 55 x 96

3

relu 1-ReLU

ReLU

55 x 55 x 96

4

norm 1-cross channel normalization with 5 channels per element

Cross Channel Normalization

55 x 55 x 96

5

pool 1

3x3 max pooling with stride [2 2] and padding [0 0 0 0]

Max Pooling

27 x 27 x 96

6

conv2-256 5 x 5 x 48 convolution with stride [11] and padding [2 2 2 2]

Convolution

27 x 27 x 96

7

relu2 -ReLU

ReLU

27 x 27 x 96

8

norm2-Cross channel normalization with 5 channels per element

Cross Channel Normalization

27 x 27 x 96

9

pool2

3x3 max pooling with stride [2 2] and padding [0 0 0 0]

Max Pooling

13 x 13x256

10

conv3-384 3 x 3 x 256 convolutions with stride [1 1] and padding [1 111)

Convolution

13 x 13x384

11

relu3-ReLU

ReLU

13 x 13x384

12

conv4 - 384 3 x 3 x 192 convolutions with stride [1 1) and padding [1 111]

Convolution

13 x 13x384

13

relu4-ReLU

ReLU

13 x 13x384

14

conv3-256 3 x 3 x 192 convolutions with stride [11] and padding [1111]

Convolution

13 x 13x256

15

relu5-ReLU

ReLU

13 x 13x256

16

poo!5

3x3 max pooling with stride [2 2] and padding [0 0 0 0]

Max Pooling

6 x 6 x 256

17

fc6

4096 fully connected layer

Fully Connected

1 x 1 x 4096

18

relu6-ReLU

ReLU

1 x 1 x 4096

19

fc7 -4096 fully connected layer

Fully Connected

1 x 1 x 4096

20

relu7-ReLU

ReLU

1 x 1 x 4096

21

special_2

64 fully connected layer

Fully Connected

1 x 1 x 64

22

Relu-ReLU

ReLU

1 x 1 x 64

23

fc8_2 -5 fully connected layer

Fully Connected

1x1x5

24

Softmax -Softmax

Softmax

1x1x5

25

Class output cross entropyex

Classification Output

-

Training and Testing Schemes

The developed CNN architecture is trained to classify images into normal and glaucoma and, further, glaucomatous cases in terms of severity using the local and public image sub-groups. From the acquired dataset, the training datasets are provided as inputs to the network. The images are then preprocessed, augmented, downsized, and fed to the CNN. One hundred images are selected for testing. The rest are selected for training the network. Many iterations are carried out randomly to generalize the performance. Stochastic gradient descent momentum training, also known as the steep descent, with a batch size of 30 samples is used to reduce the entropy loss. The momentum set is 0.9. The system is iterated with learning rates of 0.1, 0.01, 0.001, 0.0001, and 0.00001 with 300 epochs. The training rate must be optimal, as too low or high may cause long computational time or error respectively. Data augmentation is employed to prevent the problem of overfitting with the aim to artificially enlarge the training data set. Horizontal and vertical flipping is carried out randomly in training and image rotation is done via +30° to -30° [40].

Experiments

Performance Assessment

All the retinal fundus images are resized to 227x227 and subjected to the developed CNN. Experiments are performed on MATLAB, Intel Core i7 processor, 16GB RAM with GPU. The evaluation performance was quantitatively assessed using the metrics shown in Table 5.3

TABLE 5.3

Performance Parameters Used for the Developed CNN Model

Parameter

Expression

Sensitivity

Specificity

Accuracy

Precision / Positive Predictive Value

Recall

F-Score

TABLE 5.4

Results of the Developed CNN for 227x227 Input Size

Learning Rate

Sensitivity

Specificity

Accuracy

Precision

F-Score

0.1

0.95

0.966

0.94

0.95

0.84

0.01

0.9

0.941

0.956

0.91

0.88

0.001

0.98

0.984

0.989

1

1

0.0001

0.89

0.924

0.923

0.8

0.842

0.00001

0.797

0.832

0.813

0.9

0.857

Upon generalization of the results, the experiment is randomly iterated for 50 times. The average of all performance metrics for all the 50 iterations are computed and recorded. Tables 5.4 and 5.5 presents the results for the developed CNN.

The CNN can also classify the glaucomatous images into mild, moderate, and severe glaucoma. The OHT class is also added, which will serve as an indicator for glaucoma progression. Table 5.6 shows the performance of the CNN in terms of accuracy in discriminating the glaucomatous cases into different grades. Figure 5.4 Shows accuracy plot for various learning rates in grading glaucoma severity level.

Discussion and Conclusion

A CNN can have many hyper-parameters which may require further tuning, including number of fully connected layers, convolutional layers, pooling layers, filters, hidden nodes, learning rate, etc. Construction of a CNN can be time consuming if it is designed from scratch and also expensive computationally. Transfer learning could help to solve the problem by leveraging the features trained by a pre-training DL model and then applying it to various datasets. A CNN provides better performance if there is a greater number of datasets [41]. From the results, except for the learning rate 0.001, all the others have a comparatively low performance rate in terms of accuracy. The network might have missed some subtle information that serves to be vital w'hen trained with those learning rates. Image

TABLE 5.5

Average Performance of the CNN Model with 50 Iterations

Learning Rate

Sensitivity

Specificity

Accuracy

Precision

F-Score

0.1

0.956

0.932

0.921

0.94

0.84

0.01

0.965

0.949

0.938

0.9

0.852

0.001

0.949

0.951

0.9765

0.9

1

0.0001

0.872

0.919

0.88

0.75

0.857

0.00001

0.791

0.832

0.833

0.88

0.81

TABLE 5.6

Performance of CNN in Grading Glaucoma in Terms of Accuracy

Learning Rate

Accuracy

Early

Moderate

Deep

OHT

0.00001

0.65

0.6

0.58

0.75

0.0001

0.9

0.85

0.85

0.9

0.001

0.977

0.96

0.98

0.98

0.01

0.88

0.87

0.77

0.65

size also could play an important factor. Larger image size may require more computational time. To get a compromise between the two, 227x227 is chosen in the model. The model works well with different kernel sizes to attract even minor changes. Approximately 99% specificity is also achieved, making the system robust to almost all normal and abnormal cases. The experiment is iterated and the average of all the performance metrics is computed. Preprocessing is done to preserve the FOV of the images since the system can be used to further grade the disease. Table 5.7 shows the summary of and comparison of various DL methods developed for detecting glaucoma. One of the main merits of using CNN is that it does not require the normal steps of feature extraction, ranking, dimensionality reduction, etc. Feature maps are generated by extracting features automatically after each layer and taking the best features based on self-decision. It is seen that design and development of such a tool will definitely help ophthalmologist to diagnose glaucoma even at the onset of the disease or when analyzing the risk factors associated with it. The method developed is able to achieve an accuracy of

Accuracy vs. learning rate for glaucoma stages

FIGURE 5.4 Accuracy vs. learning rate for glaucoma stages.

TABLE 5.7

Summary and Comparison of Various DL Methods Developed for Detecting Glaucoma

Method

Datasets

Performance Indices

Al-Banderet al [43]

23 layer CNN. SVM

RIM-ONE

Accuracy - 85%, Sensitivity - 85%, Specificity - 89.8%

Fu et al [44]

Ensemble of 4 CNN

ORIGA. SCES

Accuracy-91.83%, Sensitivity - 84.8%, Specificity - 83.8%

Chen et al [21 ]

6 layer CNN

ORIGA. SCES

AUC - 83% to 88%

Shibata et al [45]

Transfer Learning with ResNet

Private

AUC - 96.5%

N E Benzebouchi et al [24]

Cooperative CNN

RIM-ONE

Accuracy - 96.9%, Sensitivity - 96.5%, Specificity - 97.3%

Proposed CNN

25 layer CNN

5 Public dataset, 1155 images

Accuracy - 98.9%, Sensitivity - 98%, Specificity - 98.4%, Precision and F-Score - 100%

98.9% with 98% sensitivity and specificity. The CNN can detect the normal class and glaucoma class and also can grade the images into classes: early, moderate, and deep. The OHT class is also included to assess the risk factor. When DL is used, original fundus image resolution can be reduced, taking into account the limitations of computer memory and hardware. Though there is a considerable amount of dataset, the DL method warrants much more dataset. It can be seen from [42] that a greater number of datasets is used in the study. From Table 5.2 the categorization of images may not be completely perfect due to potential bias. The developed CNN model can aid in early diagnosis of glaucoma and also predict the suspect class with a high accuracy. In future, the system can be developed with multiscale architectures to grade glaucoma with a greater number of images so that the computational time gets reduced further.

Conflict of Interest

The authors disclose no potential conflicts of interest and funding.

REFERENCES

1. A. Sommer. J.M. Tielsch, J. Katz, H.A. Quigley, J.D. Gottsch, J. Javitt. K. Singh, “Relationship between intraocular pressure and primary open angle glaucoma among white and black Americans”, The Baltimore Eye Survey. Archives of Ophthalmology 109(8) (1991). pp. 1090-1095.

  • 2. J. Nayak, U.R. Acharya, P.S. Bhat, A. Shetty, T.C. Lim, “Automated diagnosis of glaucoma using digital fundus images”, Journal of Medical Systems 33(5) (2009). pp. 337-346.
  • 3. P.S Mittapalli, G.B. Kande, “Segmentation of optic disc and optic cup from digital fundus images for the assessment of glaucoma". Biomedical Signal Processing and Control 24 (2016), pp. 24-36.
  • 4. S. Miglior, M. Guareschi, E. Albe, S. Gomarasca, M. Vavassori, N. Orzalesi, “Detection of glaucomatous visual field changes using the Moorfields regression analysis of the Heidelberg retina tomography”. American Journal of Ophthalmology 136(1) (2003), pp. 26-33.
  • 5. P. Betz, F. Camps. J. Collignon-Brach. G. Lavergne, R. Weekers, “Biometric study of the disc cup in open-angle glaucoma”. Graefe’s Archive for Clinical and Experimental Ophthalmology 218 (2) (1982), pp. 70-74.
  • 6. R. Chrastek et al. “Automated segmentation of the optic nerve head for diagnosis of glaucoma". Medical Image Analysis 9(4) (2005), pp. 297-314.
  • 7. The Eye Digest. “Eye exam and tests for glaucoma diagnosis”, The University of Illinois Eye and Ear Infirmary, Archived 2012.
  • 8. Garway-Heath et al. “Vertical cup-disc ratio in relation to optic disc size: its value in the assessment of glaucoma suspect”. British Journal of Ophthalmology 82(10) (1998), pp. 1118-1124.
  • 9. Y. Lakshmanan, R.J. George, “Stereo acuity in mild, moderate and severe glaucoma”. Ophthalmic and Physiological Optics 33(2) (2013), pp. 172-178.
  • 10. A.J. Vingrys, "The many faces of glaucomatous optic neuropathy”. Clinical and Experimental Optometry, 83 (2000), pp. 145-160.
  • 11. R. Sihota, D. Angmo. D. Ramaswamy, T. Dada, “Simplifying ‘target’ intraocular pressure for different stages of primary open-angle glaucoma and primary angle- closure glaucoma”, Indian Journal of Ophthalmology 66 (2018), pp. 495-505.
  • 12. M.L. Giger, “Computer-aided diagnosis", RSNA Categorical Course in Physics (1993), pp. 283-298.
  • 13. H. Takahashi et al, “Applying artificial intelligence to disease staging: Deep learning for improved staging of diabetic retinopathy”, PLoS One 12(6) (2017).
  • 14. V.B. Ciesielski et al. “Applications-oriented Al research: Medicine technical report", Stanford University CA, Department of Computer Science (1979).
  • 15. C.A. Kulikowski, S.M. Weiss, “Representation of expert knowledge for consultation: CASNET and EXPERT Projects”, In P. Szolovits (Ed), Artificial Intelligence in Medicine, Chapter 2. Boulder, CO: Westview Press (1982).
  • 16. http://www.glaucoma.org/glaucoma/diagnostic-tests.php.
  • 17. D. Kourkoutas et al. “Glaucoma risk assessment using a non-linear multivariable regression method”, Computer Methods and Programs in Biomedicine 108(3) (2012), pp. 1149-1159.
  • 18. A. Sevastopolsky, “Optic disc and cup segmentation methods for glaucoma detection with modification of U-Net convolutional neural networks”. Pattern Recognition and Image Analysis 27(3) (2017).
  • 19. B. Kayalibay et al. “CNN-based segmentation of medical imaging data.” CoRRubs/ 1701.03056 (2017).
  • 20. A. Geitgey, “Deep learning and convolutional neural networks.” Jun 13, 2016. [Online] Available: https://medium.eom/@ageitgey/machine-learning
  • 21. Xiangyu Chen et al, “Glaucoma detection based on deep convolutional neural network", Proceedings of the 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), IEEE (2015), pp. 715-718.
  • 22. J. Orlando, E. Prokofyeva, M. del Fresno, M.B. Blaschko, “Convolutional neural network transfer for automated glaucoma identification”, Proc. SPIE 10160, 12th International Symposium on Medical Information Processing and Analysis, 101600U
  • (2017) .
  • 23. M.V. Ferreiraa et al, “Convolutional neural network and texture descriptor-based automatic detection and diagnosis of glaucoma”, Expert Systems with Applications 110(15)
  • (2018) , pp. 250-263.
  • 24. N.E. Benzebouchi. N. Azizi. S.E. Bouziane, “Glaucoma diagnosis using cooperative convolutional neural networks”. Proceedings of ISER 88th International Conference, Venice. Italy (2017).
  • 25. R. Miikkulainen. et al. “Evolving deep neural networks”, CoRR, abs/1703.00548 (2017).
  • 26. Sun, Y„ Xue. B.. and Zhang. M. “Evolving deep convolutional neural networks for image classification”, CoRR. abs/1710.10741 (2017).
  • 27. Xie, L. and Yuille, A.L., “Genetic CNN”, CoRR. abs/1703.01513 (2017).
  • 28. Kwokleung Chan. Те-Won Lee, P. A. Sample. M. Goldbaum, R. N. Weinreb, T. J. Sejnowski, “Comparison of machine learning and traditional classifiers in glaucoma diagnosis”. IEEE Transactions on Biomedical Engineering 49(9) (2002). doi: 10.1109/ TBME.2002.802012.
  • 29. O. Sheeba. “Glaucoma detection using artificial neural network”. IACSIT, Ed. IACSIT, 6(2) (2014)
  • 30. H. Eckerson, “Deep learning past, present, and future,” 2017. [Online] Available: https://www.kdnuggets.corn/2017/05/deep-learning-big-deal.html
  • 31. A. Allam, A. Youssif, A. Ghalwash, “Automatic segmentation of optic disc in eye fundus images: A survey, ELCVIA: Electron". Electronic Letters on Computer Vision and Image Analysis 14(2015), pp. 1-20.
  • 32. A. Almazroa et al, “Optic disc and optic cup segmentation methodologies for glaucoma image detection: a survey”, Journal of Ophthalmology 2015 (2015), p. 28.
  • 33. N. Thakur, M. Juneja, “Comparative analysis on optic cup and disc segmentation for glaucoma diagnosis”, Springer Proceedings of 3rd International Conference on Advanced Computing, Networking and Informatics (2016), pp. 219-223.
  • 34. D. Calkins. “Glaucoma”, Tech. Rep. 88-207450.2016. [Online] Available: https://www. glaucorna.org/GRF_Understanding_Glaucorna_ES.pdf
  • 35. К. He, X. Zhang, S. Ren. J. Sun, “Deep residual learning for image recognition”, In CVPR (2016).
  • 36. S. Mukhopadhyay et al. “Multiscale morphological segmentation of gray scale images”, IEEE Transactions on Image Processing 12 (2003), pp. 533-549.
  • 37. S.C. Wong. A. Gatt, V. Starnatescu, M.D. McDonnell, “Understanding data augmentation for classification: When to warp?” International Conference on Digital Image ComputingTechApplDICTA2016(2016). https://doi.org/10.1109/DICTA.2016.7797091
  • 38. G. Litjens, T. Kooi, B. Bejnordi, A. Setio. F. Ciompi. M. Ghafoorian. J. Laak, B. Ginneken. C. Sanchez. “A survey on deep learning in medical image analysis”, Medical Image Analysis 42 (2017), pp. 60-68.
  • 39. U. Rajendra Acharya et al. “Automated detection of arrhythmias using different intervals of tachycardia ECG segments with convolutional neural network”. Information Sciences, 405 (2017), pp. 81-90.
  • 40. L. Wang et al. “2D/3D rigid registration by integrating intensity distance”. Optics and Precision Engineering. 22 (2014), pp. 2815-2824.
  • 41. U Rajendra Acharya et al, “Automated identification of shockable and non- shockable life threatening ventricular arrhythmias using convolutional neural networks”. Future Generation Computer Systems 79 (2017), 952-959.
  • 42. A. Krizhevsky, I. Sutskever, G. E. Hinton. “ImageNet classification with deep convolutional neural networks”, Communications of ACM 60(6) (2017). pp. 84-90.
  • 43. B. Al-Bander. W. Al-Nuaimy, M.A. Al-Taee, Y. Zheng, “Automated glaucoma diagnosis using deep learning approach”. Proceedings of 14th International Multi-Conference on Systems, Signals Devices (SSD) (2017), pp. 207-210.
  • 44. H. Fu, J. Cheng, Y. Xu, C. Zhang, D.W.K. Wong, J. Liu. and X. Cao, “Disc-Aware ensemble network for glaucoma screening from fundus image”, IEEE Transactions on Medical Imaging 37(11) (2018), 2493-2501.
  • 45. N. Shibata. M. Tanito. K. Mitsuhashi, Y. Fujino. M. Matsuura, H. Murata. and R. Asaoka, “Development of a deep residual learning algorithm to screen for glaucoma from fundus photography”, Scientific Reports 8(1) (2018), 14665.
 
Source
< Prev   CONTENTS   Source   Next >