CSE 4000: Thesis/ Project

Detection of Diabetic Retinopathy from Retinal Image

By
Md Abu Saeed
Roll:1907057

Supervisor:
Kazi Saeed Alam
Assistant Professor
Deapartment of CSE
khulna university of engineering & technology

Department of Computer Science and Engineering

Khulna University of Engineering & Technology
Khulna 9203, Bangladesh
October, 2024

1
 Contents
1 Introduction 4

2 Related work 5

3 Methodology 7
3.1 Dataset Selection . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Image Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2.1 Extract the vessels . . . . . . . . . . . . . . . . . . . . . 8
3.2.2 Extract the microaneurysm . . . . . . . . . . . . . . . . . 10
3.2.3 Merge previous results . . . . . . . . . . . . . . . . . . . 11
3.3 Data Augmentation . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.4 Model Training . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4 Current Result 13
4.1 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

5 Conclusion 15

6 References 16

2
 List of Figures
1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Vessels extraction steps . . . . . . . . . . . . . . . . . . . . . . . 9
3 Microaneurysm extraction steps . . . . . . . . . . . . . . . . . . 10
4 Final merged image . . . . . . . . . . . . . . . . . . . . . . . . . 11
5 Data augmentation . . . . . . . . . . . . . . . . . . . . . . . . . 12

3
1 Introduction
Diabetic retinopathy (DR) is a severe eye condition that arises as a complication
of diabetes, resulting from prolonged high blood sugar levels. It leads to
progressive damage to the blood vessels in the retina, which may cause vision
impairment and, if left untreated, permanent blindness. The retina is a critical part
of the eye that senses light and sends visual signals to the brain, and damage to its
blood vessels severely affects vision. DR progresses through various stages,
starting from mild, non-proliferative forms where blood vessels may leak or swell
(Non-Proliferative Diabetic Retinopathy, NPDR) to more advanced stages
marked by abnormal blood vessel growth (Proliferative Diabetic Retinopathy,
PDR). In its early stages, diabetic retinopathy is often asymptomatic, making
timely detection and treatment difficult. However, as the disease progresses,
patients may experience blurred vision, floaters, and eventual vision loss

Despite the high prevalence of diabetic retinopathy, especially in diabetic

populations, current diagnostic methods rely heavily on manual examination by
ophthalmologists using fundus photography. This process is both
time-consuming and prone to human error, particularly when detecting subtle
retinal abnormalities at early stages. With the increasing number of diabetes cases
globally, there is a growing need for automated systems that can reliably and
efficiently detect diabetic retinopathy from retinal images. Developing a machine
learning-based detection system could provide a scalable solution to aid early
diagnosis, improving outcomes by allowing for timely intervention.

4
2 Related work
In recent years, the application of convolutional neural networks (CNNs) to
medical image analysis has gained significant traction, particularly in the
detection of diabetic retinopathy (DR) [1]. A notable study focused on
developing a computer-assisted tool using CNNs for rapid and accurate
classification of retinal images to diagnose diabetic retinopathy [2]. The tool
identifies key features such as exudates, microaneurysms, and hemorrhages from
retinal images, achieving a specificity of 73.65% and an accuracy of 73.68% in
validation tests on 35,126 images from the EyePACS dataset. This approach
demonstrates the potential of CNNs in reducing the time and effort required for
manual diagnosis by clinicians [3].

Several studies have explored similar deep learning approaches to automate the
diagnosis of DR. One paper employed a deep neural network, specifically a CNN,
to classify retinal images into diabetic and non-diabetic categories, achieving over
85% accuracy during training [4], [5]. This significant improvement over
traditional algorithms highlights the efficiency of CNNs in feature extraction,
leveraging layers like ReLU activation and MaxPooling to identify key patterns in
the images [6], [7]. Another study focused on screening for non-proliferative
diabetic retinopathy (NPDR) using a combination of advanced image processing
techniques such as recursive region growing segmentation and the Moat Operator.
This automated system demonstrated high sensitivity and specificity, comparable
to the performance of experienced ophthalmologists [8], [9].

Furthermore, deep CNNs have proven to outperform traditional feature-based

classifiers, as demonstrated in research that utilized color fundus images for the
automatic detection of diabetic retinopathy. By integrating data augmentation
techniques, these CNN-based approaches achieved an impressive accuracy of
84.5%, surpassing classical methods [10]. Another noteworthy study developed a
CNN-based screening method for diabetic retinopathy using Theano and

5
achieved around 85% accuracy in two-class classification, further showcasing the
effectiveness of CNNs in medical imaging [10].

The VGGNet [11] model has also been effectively employed in the automatic
classification of DR. In one study, CNNs were used to extract features such as
blood vessels and microaneurysms, leading to an accuracy of 95.41% in detecting
the presence of diabetic retinopathy. The high accuracy of this model underscores
the importance of early detection through automated systems, which can play a
crucial role in preventing severe vision loss [3] [6].

In addition to detecting lesions, several methods have been proposed for

identifying blood vessels in retinal images, a critical step in diagnosing ocular
diseases. Traditional edge detection techniques often fail due to the poor contrast
in retinal images. However, novel operators based on the optical and spatial
properties of blood vessels, combined with matched filter detection, have shown
promising results in accurately detecting linear segments of vessels [1].

6
3 Methodology
The full methodology can be described as follows:

1. Select a dataset from Kaggle.

2. Preprocess the images to highlight features.

(a) Extract the vessels.

(b) Extract the micro-aneurysms.
(c) Merge those two images.

3. Data Augmentation

(a) A new image is created by adjusting the brightness by adding a random

delta value then clipping.
(b) Then 6 variations are made: original, brightness adjusted, and rest 4 are
90- and 180-degree rotated version of two images.

4. Then different type of models is used to classify.

(a) Using custom model having few layers.

(b) Using VGG16 with some layers.

Figure 1: Methodology

7
3.1 Dataset Selection

The Kaggle Diabetic Retinopathy Detection dataset was used, consisting of

high-resolution retinal fundus images labeled on a scale of 0 to 4, representing
different stages of diabetic retinopathy. The large size and variety of images make
it ideal for deep learning applications.

3.2 Image Pre-processing

3.2.1 Extract the vessels

1. Extract the green channel.

2. Performs histogram equalization to enhance the contrast.

3. Remove background artifacts by replacing zero pixels with mean intensity.

4. Top hat filter is applied to emphasize the vessel like structures.

5. Otsu threshold is applied to make binary image.

6. Remove small non-vessel structure length less than 100 pixels.

7. Median blur is used to remove noise.

8. Thin vessels are extracted using a matched filter technique.

9. Gaps are filled by closing.

10. Small thin vessels are removed having length less than 30 pixels.

11. Then the border circle is removed.

8
Figure 2: Vessels extraction steps

9
3.2.2 Extract the microaneurysm

1. Extract the green channel.

2. A disk-shaped structuring element is created using radius 1.

3. Then image is dilated then eroded.

4. Gaussian filtering is used to blur the image.

5. Thresholding is applied to select the top 5% pixels.

6. Then large white areas are removed having length greater than 20 pixels.

Figure 3: Microaneurysm extraction steps

10
3.2.3 Merge previous results

Both the images are merged by taking the maximum intensity at each pixel.

Figure 4: Final merged image

11
3.3 Data Augmentation

Data augmentation is employed to increase the size of the training dataset and
improve the generalization ability of the model. The augmentation process
involves creating new variations of each image by modifying its brightness and
applying rotations. The steps are as follows:

1. Brightness Adjustment: A random delta value is added to the brightness of

the image, and the resulting pixel values are clipped to ensure they remain
within a valid range. This step simulates different lighting conditions that
might be present during real-world imaging.

2. Rotation: Six variations of each image are generated. These include: The
original image. The brightness-adjusted image. 90-degree and 180-degree
rotated versions of both the original and brightness-adjusted images.

This augmentation strategy ensures that the model is trained on a diverse set of
images, improving its robustness to variations in lighting and orientation.

Figure 5: Data augmentation

12
3.4 Model Training

1. Custom CNN: A custom model with a few convolutional layers and

MaxPooling was designed to learn from the extracted features.

2. VGG16 Transfer Learning: The VGG16 architecture was used with

pre-trained layers for feature extraction, and the final layers were
customized for DR classification.

4 Current Result
Model accuracy on validation set is around 70%. (unstable)

13
4.1 Future work

I am planning to do these to complete my thesis:

1. Fix the model issue to make it stable.

2. Combining Additional Preprocessing Techniques such as noise reduction

filters, image sharpening to further enhance the visibility of key features like
microaneurysms and blood vessels.

3. Applying Transfer Learning models like ResNet or Inception, along with

VGG16, to improve classification accuracy and handle more complex retinal
features.

4. Expanding the Dataset to include more diverse and larger datasets, such as
those with multi-modal imaging techniques, to improve the generalization of
the model across different populations.

14
5 Conclusion
This thesis focuses on developing a model for classifying diabetic retinopathy
using convolutional neural networks (CNNs) and advanced image processing
techniques. The methodology includes preprocessing techniques to enhance
retinal features, feature extraction to isolate blood vessels and microaneurysms,
and data augmentation to create robust training datasets. Both custom CNN
models and transfer learning with VGG16 have been employed to get good result.
Since the system demonstrates unstable performance in classifying retinal
images, future work will focus on to resolve it, combining additional
preprocessing techniques to further enhance feature visibility and applying more
advanced transfer learning models to improve accuracy. The ongoing expansion
of the dataset and exploration of ensemble methods will further refine the model,
ultimately aiming for a more accurate and scalable solution for diabetic
retinopathy detection.

15
6 References

References
[1] C. Sinthanayothin, J. F. Boyce, T. H. Williamson, H. L. Cook, E. Mensah,
S. Lal, and D. Usher, “Automated detection of diabetic retinopathy on digital
fundus images,” Diabetic medicine, vol. 19, no. 2, pp. 105–112, 2002.

[2] R. Ghosh, K. Ghosh, and S. Maitra, “Automatic detection and classification

of diabetic retinopathy stages using cnn,” in 2017 4th International
conference on signal processing and integrated networks (SPIN). IEEE,
2017, pp. 550–554.

[3] G. Garcı́a, J. Gallardo, A. Mauricio, J. López, and C. Del Carpio, “Detection

of diabetic retinopathy based on a convolutional neural network using retinal
fundus images,” in Artificial Neural Networks and Machine
Learning–ICANN 2017: 26th International Conference on Artificial Neural
Networks, Alghero, Italy, September 11-14, 2017, Proceedings, Part II 26.
Springer, 2017, pp. 635–642.

[4] H. Pratt, F. Coenen, D. M. Broadbent, S. P. Harding, and Y. Zheng,

“Convolutional neural networks for diabetic retinopathy,” Procedia
computer science, vol. 90, pp. 200–205, 2016.

[5] M. Raju, V. Pagidimarri, R. Barreto, A. Kadam, V. Kasivajjala, and

A. Aswath, “Development of a deep learning algorithm for automatic
diagnosis of diabetic retinopathy,” in MEDINFO 2017: precision healthcare
through informatics. IOS press, 2017, pp. 559–563.

[6] D. S. Sisodia, S. Nair, and P. Khobragade, “Diabetic retinal fundus images:

Preprocessing and feature extraction for early detection of diabetic
retinopathy,” Biomedical and Pharmacology Journal, vol. 10, no. 2, pp.
615–626, 2017.

[7] M. A. H. Raj, M. Al Mamun, and M. F. Faruk, “Cnn based diabetic

16
 retinopathy status prediction using fundus images,” in 2020 IEEE region 10
symposium (TENSYMP). IEEE, 2020, pp. 190–193.

[8] M. Ragab, A. S. A.-M. Al-Ghamdi, B. Fakieh, H. Choudhry, R. F. Mansour,

and D. Koundal, “Prediction of diabetes through retinal images using deep
neural network,” Computational Intelligence and Neuroscience, vol. 2022,
no. 1, p. 7887908, 2022.

[9] S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and M. Goldbaum,

“Detection of blood vessels in retinal images using two-dimensional
matched filters,” IEEE Transactions on medical imaging, vol. 8, no. 3, pp.
263–269, 1989.

[10] K. Xu, D. Feng, and H. Mi, “Deep convolutional neural network-based early
automated detection of diabetic retinopathy using fundus image,”
Molecules, vol. 22, no. 12, p. 2054, 2017.

[11] C. Alippi, S. Disabato, and M. Roveri, “Moving convolutional neural

networks to embedded systems: the alexnet and vgg-16 case,” in 2018 17th
ACM/IEEE International Conference on Information Processing in Sensor
Networks (IPSN). IEEE, 2018, pp. 212–223.

17