0% found this document useful (0 votes)
30 views18 pages

Proliferative Diabetic Retinopathy Diagnosis Using Varying Scales Filter Banks and Double Layered Thresholding

Uploaded by

Anum Abdul Salam
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
30 views18 pages

Proliferative Diabetic Retinopathy Diagnosis Using Varying Scales Filter Banks and Double Layered Thresholding

Uploaded by

Anum Abdul Salam
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 18

diagnostics

Article
Proliferative Diabetic Retinopathy Diagnosis Using
Varying-Scales Filter Banks and Double-Layered Thresholding
Noor ul Huda 1 , Anum Abdul Salam 2 , Norah Saleh Alghamdi 3, * , Jahan Zeb 2
and Muhammad Usman Akram 2, *

1 Center for Advanced Studies in Telecommunications (CAST), COMSATS Institute of Information Technology,
Islamabad 45550, Pakistan; [email protected]
2 Computer and Software Engineering Department, College of Electrical and Mechanical Engineering ,
National University of Sciences and Technology, Islamabad 24090, Pakistan;
[email protected] (A.A.S.); [email protected] (J.Z.)
3 Department of Computer Sciences, College of Computer and Information Sciences,
Princess Nourah bint Abdulrahman University, P.O. Box 84428, Riyadh 11671, Saudi Arabia
* Correspondence: [email protected] (N.S.A.); [email protected] (M.U.A.)

Abstract: Diabetic retinopathy is one of the abnormalities of the retina in which a diabetic patient
suffers from severe vision loss due to an affected retina. Proliferative diabetic retinopathy (PDR) is the
final and most critical stage of diabetic retinopathy. Abnormal and fragile blood vessels start to grow
on the surface of the retina at this stage. It causes retinal detachment, which may lead to complete
blindness in severe cases. In this paper, a novel method is proposed for the detection and grading of
neovascularization. The proposed system first performs pre-processing on input retinal images to
enhance the vascular pattern, followed by blood vessel segmentation and optic disc localization. Then
various features are tested on the candidate regions with different thresholds. In this way, positive
and negative advanced diabetic retinopathy cases are separated. Optic disc coordinates are applied
for the grading of neovascularization as NVD or NVE. The proposed algorithm improves the quality
of automated diagnostic systems by eliminating normal blood vessels and exudates that might cause
hindrances in accurate disease detection, thus resulting in more accurate detection of abnormal blood
vessels. The evaluation of the proposed system has been carried out using performance parameters
Citation: Huda, N.u.; Salam, A.A.;
such as sensitivity, specificity, accuracy, and positive predictive value (PPV) on a publicly available
Alghamdi, N.S.; Zeb, J.; Akram, M.U.
standard retinal image database and one of the locally available databases. The proposed algorithm
Proliferative Diabetic Retinopathy
gives an accuracy of 98.5% and PPV of 99.8% on MESSIDOR and an accuracy of 96.5% and PPV of
Diagnosis Using Varying-Scales Filter
Banks and Double-Layered
100% on the local database.
Thresholding. Diagnostics 2023, 1, 0.
https://doi.org/ Keywords: image processing; biomedical imaging; diabetic retinopathy;
proliferative diabetic retinopathy; autonomous disease detection; fundus image analysis
Academic Editor: Firstname Lastname

Received: 2 June 2023


Revised: 9 June 2023 1. Introduction
Accepted: 10 June 2023 Diabetic retinopathy (DR) is a vascular complication of eye. It is the most severe
Published:
abnormality among all diabetic eye diseases [1]. In DR, at first a lesion starts to appear
on the retina, and then it results in the bleeding of blood vessels and capillaries on the
surface of the retina. Due to the leakage of blood vessels, the oxygen supply to the retina
Copyright: © 2023 by the authors.
decreases and as a consequence the brain stimulates the formation of new blood vessels to
Licensee MDPI, Basel, Switzerland. fulfill the requirement of oxygen. These new vessels may bleed and cause the detachment
This article is an open access article of the retina and ultimately loss of vision [2,3]. Out of all DR patients, few suffer from
distributed under the terms and proliferative diabetic retinopathy (PDR), but if it is not diagnosed in a timely manner, the
conditions of the Creative Commons disease may cause severe destruction. Neovascularization appears as a tortuous collection
Attribution (CC BY) license (https:// of blood vessels and is quite destructive because these vessels grow abnormally out of
creativecommons.org/licenses/by/ the retina into the clear vitreous gel [4,5]. Therefore, vessels grow beyond the supporting
4.0/). structure of the retina and they are very disposed to bleeding, particularly when they

Diagnostics 2023, 1, 0. https://doi.org/10.3390/diagnostics1010000 https://www.mdpi.com/journal/diagnostics


Diagnostics 2023, 1, 0 2 of 18

arise near the optic disc. Even a small rise in blood pressure can cause hemorrhages in
this case. If bleeding appears in the vitreous humor it can affect the visual system. If this
bleeding becomes extensive, it results in a painless and rapid blackening of the vision.
Neovascularization [6] is divided into two types [7]:
• Neovascularization on disc (NVD): If the new vessel formation occurs within one disc
diameter of the optical disc then this is categorized as NVD or neovascularization on disc
• Neovascularization elsewhere (NVE): If new vessel formation occurs elsewhere on the
surface of the retina, then this is called neovascularization elsewhere (NVE).
Figure 1 highlights the classification of neovascularization. The article consists of five
sections. Section 2 of the paper describes the existing work and the main contributions
presented by the proposed method. The details of the proposed system are given in
Section 3. Section 4 shows the results and evaluation while Section 5 is composed of the
conclusion and future work.

Figure 1. Digital fundus image with NVD and NVE.

2. Related Work
Several techniques have been proposed so far for the automated screening and detec-
tion of non-proliferative diabetic retinopathy and its stages [8–10], However, for detection
of PDR a very few work has been carried out. We have organized the related literature
based on two subcategories, i.e., PDR detection using machine learning techniques and
PDR detection using deep learning techniques.

2.1. Machine-Learning-Based Algorithms


Gotman et al. [11] proposed a method for the detection of new vessels on the disc using
a support vector machine (SVM) classifier. They used a watershed transform combined
with 2D Gaussian for the segmentation of blood vessels. Fifteen features were used for
the classification of the SVM classifier and they achieved an area under receiver operating
characteristics curve (ROC) of 0.909. They came up with a feature set to detect neovascu-
larization but they limited their scope to NVD only. An amplitude modulation–frequency
modulation (AM-FM)-based method was proposed by Agurto et al. [12]. They followed a
top-down approach for the detection of NVD. They came up with a sound set of features
and used K-means clustering on these feature sets. Jelinek et al. [13] proposed a new
technique for the detection of proliferative diabetic retinopathy from angiograms. They
used 27 labeled images and achieved an accuracy of 0.90 with the selection of six features.
Derivatives of the Gaussian wavelet were used in their work for the segmentation of blood
vessels. Mudigonda et al. [14] proposed a method for the detection of neovascularization in
retinal fundus images using fractal analysis. The proposed technique used colored fundus
images as the input, followed by ROI extraction, i.e., extraction of the region around the
optic disc. The green channel was extracted from the ROI region to obtain the maximum
Diagnostics 2023, 1, 0 3 of 18

information from the vessels. The vessels were extracted using a Gabor filter, and the result-
ing magnitude image was converted to a binary image. The resultant image was analyzed
using a fractal analysis box-counting method that identified vessel bundles near the optic
disc region, i.e., neovascularization in the optic disc region (NVD). Among ten images,
five with neovascularization had a fractal mean value of 1.66 and five images with no
neovascularization in the optic disc region resulted in a fractal mean value of 1.58. Saranya
et al. [15] used the fuzzy C-means (FCM) technique for blood vessel segmentation. They
used a set of features that included the gradient, gradient variation, gray-level coefficient of
variation, moment invariant-based features, and tortuosity for a k-nearest neighbor (KNN)
classifier. They achieved an accuracy of 96.5% on the DRIVE and MESSIDOR datasets. A
research methodology for the automated detection of neovascularization for PDR proposed
by Sohini Roy Chowdhury et al. [16] describes a technique to detect neovascularization
from fundus images and classify it as neovascularization in the optic disc region (NVD) or
neovascularization elsewhere (NVE). The green plane is extracted from the input image
and normalized in the [0 1] intensity range. The region of interest (ROI) is extracted for
both types of neovascularization, leading to vessel detection from both ROIs. Textural,
structural, and intensity-based features are used to classify NVD and NVE. The proposed
method was trained and tested on 40 images (30 normal, ten with PDR) from the STARE
database and 17 images from a local dataset. Accuracies of 87.6% and 92.1% were obtained
for NVD and NVE, respectively.
Lee et al. [17] proposed a new vessel-detection method that includes statistical texture
analysis (STA), high-order spectrum analysis (HOS), and fractional analysis. They used a
total of 137 images in their work and achieved an area under the curve of 99.3%. A method
based on the following-the-line approach for the segmentation of vessels was proposed by
R.A. Welikala et al. [18]. They used two-line approaches and two different sets of features
for the retraction of true abnormal blood vessels. They used 60 images from the MESSIDOR
dataset for evaluation purposes and classified these images on the basis of an SVM classifier.
They achieved an area under the curve (AUC) value of 0.96. Shuang Yu et al. [19] proposed
a novel technique for the automation of neovascularization in the optic disc region (NVD).
A fundus image is pre-processed followed by the application of a Gabor filter to extract
blood vessels. Twenty-one texture-based and vessel-based features are extracted to classify
an image as normal or NVD using support vector machines (SVM). Sixty-six retinal images
(15 NVD, 50 normal) were extracted from the globally available MESSIDOR, HRF, and
DIARETDB0 datasets to test and train the proposed technique. A sensitivity of 15/16 and
specificity of 47/50 were achieved using the proposed methodology.
In 2016 research was conducted by Diego F. G. Coelho [20] that aimed to detect
NVD from fundus images. In the proposed methodology, fundus images are analyzed by
calculating the gradient magnitude of the Fourier power spectrum followed by extraction
of the angular spread. Entropy and spatial variance are used to categorize an image as a
normal image or image with NVD using a linear statistical classifier. An accuracy of 100%
was achieved when the proposed technique was tested on ten images (five normal, five
NVD) extracted from the MESSIDOR database. Akram et al. [7] presented a method for the
detection of PDR. Their proposed method extracted a number of features based on vascular
patterns for the proper representation of normal and abnormal vessels. A modified m-
mediods-based classifier was used for the proper discrimination of abnormal vessels from
normal ones. Another machine-learning-based technique for automated NVD detection
was proposed by Shuang Yu et al. [21]. The proposed algorithm takes a fundus image and
extracts the ROI, i.e., the disc region. Vessels are extracted using multilevel Gabor filters. A
feature vector with 42 features (morphological and texture based) is extracted from both
normal and NVD images, followed by a reduction in size to eighteen. A reduced feature
vector was used to train and test 424 (134 NVD, 290 non-NVD) retinal fundus images. An
accuracy of 95.23% was observed.
Christodoulidis et al. [22] proposed a novel technique for the detection of NVD from
fundus images. The proposed research states that NVD detection from retinal fundus
Diagnostics 2023, 1, 0 4 of 18

images can be improved by adding a second-order statistical feature to the existing feature
set containing structural, vessel-based, and intensity-based features. The image is pre-
processed, followed by vessel detection. The vessel junctions are extracted by applying the
Tensor voting technique, which highlights the local maxima, indicating the junctions of
vessels. The suggested feature addition to the feature vector improved the sensitivity to 0.84.
Mona Leeza et al. [23] proposed an algorithm in 2019 to detect the severity level of diabetic
retinopathy using the bag-of-features approach. The algorithm is composed of five phases
starting from local feature extraction from retinal images using SURF. K-means clustering
is used to cluster the extracted features for dictionary generation. The algorithm proceeds
by max pooling to accumulate features followed by the construction of histograms of
oriented gradient (HoG). SVM and artificial neural networks are used to classify the retinal
image as normal, mild NPDR, moderate NPDR, severe NPDR, and PDR. The algorithm
resulted in 95.92% and 98.90% sensitivity and specificity, respectively. Research conducted
by Lei Zhang et al. [24] described another algorithm to screen for PDR using a modified
matched-filter approach. In the proposed technique, the result from Gaussian is proceeded
by subtraction of the mean to eliminate the false positives that occur due to step edge noise.
The accuracy of the algorithm was evaluated on the ZUEYE database, which resulted in an
accuracy of 95%.

2.2. Deep-Learning-Based Algorithms


A technique was proposed in 2019 [25] to detect and categorize diabetic retinopathy
using a deep convolution neural network (CNN) of five layers. The algorithm starts
with pre-processing of the images, followed by an ensemble CNN model. The ensemble
CNN model is composed of five deep CNN models, i.e., Resnet50, Inceptionv3, Xception,
Dense121, and Dense169. The CNN model classifies the input image as normal, mild,
moderate, severe, or PDR. The algorithm was tested on the Kaggle dataset, composed of
35126 colored retinal images. Sixty-four percent of images in the dataset were used for
training, 20% images were used for testing, and 16% were used for validation. Specificities
of 0.40, 0.99, 0.95, 0.98, and 0.99 were observed for each category, respectively. In 2022 [26],
a neural-network-based framework was proposed using optical coherence tomography
(OCT) images. The proposed algorithm classified an OCT image as normal or diseased
using 3D feature extraction. Initially, segmentation was performed to extract 12 layers from
the input image followed by feature extraction, i.e., thickness and angle calculation. The
extracted features were then fused and passed to the neural network to make a decision,
yielding an accuracy of 96.61%.
Ayesha et al. [27] proposed three deep neural frameworks for diabetic retinopathy
grading using retinal fundus images. The first framework used cascaded architecture to
grade a retinal image among five grades of PDR using a three-layer CNN architecture.
The second framework utilized the hue saturation value (HSV), red green blue (RGB),
and normalized input image to apply ensemble-based architecture, where the final results
were deduced using average pooling from each CNN model. The third framework incor-
porated a long short-term memory (LSTM) module to enhance the network memorizing
capabilities. The EyePACS dataset containing 88,702 retinal images was used to train and
test the proposed framework. Among all, the ensemble-based architecture outperformed,
resulting in an accuracy of 83.78%. Another framework proposed by Tang et al. [28] seg-
mented and localized neovascularization using a deep learning architecture. The proposed
algorithm starts with image pre-processing followed by dividing the input image into
non-overlapping patches. To train the neural network, the ground truth containing neo-
vascularization in each patch was passed as a training dataset, which classified each pixel
of the patch as neo or non-neo. The dataset was divided into validation, training, and
testing sets, yielding an accuracy of 0.9948 on a dataset with 50 images. This research was
further extended [29] using transfer learning on pre-trained models that included AlexNet,
GoogLeNet, ResNet18, and ResNet50 pre-trained on ImageNet. Ground truth patches
were used for training these models, followed by testing the models. Another module
Diagnostics 2023, 1, 0 5 of 18

utilized the pre-trained model for feature extraction followed by classification using SVM.
In addition to using pre-trained models separately, a combination of ResNet and GoogleNet
was proposed that yielded the highest accuracy of 0.9157.
Another algorithm [30] utilized ResNet to detect retinal neovascularization from
retinal fundus images. The proposed framework pre-processed the input image to enhance
the contrast and remove noise followed by training ResNet. The proposed model was
trained on 3662 retinal images from a local dataset containing healthy images (Label 0),
neovascularized images (Label 2), and diabetic retinopathy images (Label 1). Due to the
residual properties of ResNet, the model resulted in an accuracy of 0.88 on 1992 retinal
fundus images. A review conducted by Salamat et al. [6] summarized and comprehended
66 papers aiming to detect and classify diabetic retinopathy using various techniques.
The paper presented the past 8 years of research articles starting from 2019, indicating the
techniques used, the dataset, and the evaluation metrics along with the type of classification
or number of classes. The majority of the techniques presented in the review classified an
image into healthy or diseased without classifying it by the degree/severity of disease.
Our proposed method is an extension of [7]. Our method grades the PDR with-
out the use of a classifier and utilizes a simple feature-extraction approach to minimize
false-positive detection. The main issues with most of the algorithms mentioned are the
appearance of false-positive regions for abnormal blood vessels and evaluation on a very
small dataset. The proposed algorithm is novel in the sense that it addresses both NVD and
NVE. It uses a very robust method for the detection of optic disc coordinates. It gives al-
most no false detection, unlike other cases. The algorithm extracts optic disc and candidate
abnormal blood vessel regions using the vascular structure and filter bank, respectively.
Double-layered thresholds on the basis of pattern analysis for the detection of PDR are
applied, which grade the input retinal image as NVD or NVE.

3. Materials and Methods


The proposed system follows the following steps. It first performs pre-processing on
the images to enhance the vascular pattern, which is followed by blood vessel segmentation
and optic disc localization. Various features are tested on the candidate regions with differ-
ent thresholds for the separation of positive and negative advanced diabetic retinopathy
cases. Optic disc coordinates are applied for the grading of neovascularization as NVD or
NVE. The algorithms improve the quality of the automated system by eliminating normal
blood vessels and exudates for the accurate detection of abnormal blood vessels. The
whole process can be arranged into four basic steps, which involve pre-processing, vessel
extraction, ROI processing, and post-processing. Figure 2 shows the complete flow of the
automated system for grading PDR.
Diagnostics 2023, 1, 0 6 of 18

Figure 2. Proposed flow diagram for the detection and grading of PDR.

3.1. Preprocessing
An automated assessment for pathologies of the retina initially requires the pre-
processing of a digital fundus image. An inverted green channel is used as it enhances the
vascular patterns against the dark background. All images are scaled to the same size, not
disturbing their aspect ratio. The dark background of the image is not really black as it
contains some of the lighter regions and an amount of noise. It is necessary, for the proper
extraction of the vascular pattern in the retina, to separate the noisy background from
the image. For this purpose, background segmentation is carried out. The method first
creates a binary mask for the background by using the mean and variance and then it elimi-
nates the small noisy pixel values from the background by using different morphological
operators [31]. Figure 3 shows the images after pre-processing.

Figure 3. Background segmentation, (A) Extracted Background Mask (B) Background-segmented


and scaled image in the inverted green channel.

3.2. Vessel Segmentation


After pre-processing, a 2D Gabor wavelet is applied to the image to enhance the
vascular pattern, so that the abnormal blood vessels, which are thin and less visible,
become visible and prominent [32]. Gabor wavelets can be set to specify a direction for
vessel segmentation. They are very sensitive to small edges and have directional selectivity
Diagnostics 2023, 1, 0 7 of 18

capability. They also act as a filter for the background noise. In this work, 2D continuous
wavelet Gabor transform (CWT) is used. It is defined in (1):
Z
Tψ (b, θ, a) = Cψ−1/2 a exp ( jkb)ψ̂∗ ( ar−θ k) ĝ(k)d2 k (1)

where j = −1, and ψ̂∗ and ĝ denote a Fourier transform. The 2D Gabor wavelet is defined
as:
1
ψG (x) = exp( jk0 x) exp(− |Ax|2 ) (2)
2
1
ψ̂G (x) = (det( B))1/2 exp(− ( B(k − k0 )2 )) (3)
2
where k0 ∈ R2 is a vector that defines the frequency of the complex exponential, B = A−1
 −1/2 
ϵ 0
and A = with elongation ϵ ≥ 1 is a 2 × 2 positive definite diagonal matrix
0 1
that defines the wavelet anisotropy and its elongation in any desired direction.
The Gabor wavelet transform Mψ (b, a) is computed for each pixel position and its
scale value is considered. In addition, θ spans from 0o to 165o at steps of 10o and its
maximum is taken.
Mψ (b, a) = max | Tψ (b, θ, a)| (4)
Multilayered and adaptive thresholding techniques are applied to create a binary
mask for blood vessels after the completion of blood vessel enhancement [33]. The masking
process assigns 1 to all vessel pixels and 0 to all non-vessel pixels.

3.3. Abnormal Vessel Detection


The abnormal blood vessel extraction process is explained in Figure 4. Two copies of
the image are created after pre-processing. A Gaussian blur filter of 3×3 is applied to the
first copy of the image. This filters out the minor details. This copy of the image is further
processed using a 2D Gabor filter that is set on such frequencies and directions to enhance
the normal blood vessels only. The second copy of the image is fed to a sharpening filter to
enhance the small details in the image. Then a 2D Gaussian filter is applied to this copy of
the image with set frequencies that also enhance the small details as well as normal blood
vessels. Tables 1 and 2 show the selected values for each parameter of Gabor wavelet for
blood vessel enhancement for NVD and NVE cases, respectively.
Diagnostics 2023, 1, 0 8 of 18

Figure 4. Flow for abnormal blood vessel extraction.

Table 1. Parameter values of Gabor wavelet for NVD.

Parameters Value for Normal Value for Abnormal


Blood Vessels Blood Vessels
Dilation (a) 11 1.8
Elongation (ϵ) 5 1
Rotation Angle (θ) 10° 10°
k0 [0,2.5] [0,2]

Table 2. Parameter values of Gabor wavelet for NVE.

Parameters Value for Normal Value for Abnormal


Blood Vessels Blood Vessels
Dilation(a) 7 2
Elongation (ϵ) 10 1
Rotation angle (θ) 10° 10°
k0 [0,2.5] [0,2.5]

Then multilayered thresholding is applied on both copies of the image as described


in Section 2.2. Both copies are then subtracted to come up with a region that contains
abnormal blood vessels and exudates only. This is our candidate region of interest.

3.4. Optic Disc Detection


The optic disc (OD) is a comparatively brighter region in the fundus image with a
bright yellowish color and circular shape, but it also shows some variation in brightness
and color if some disease is present, which can make OD detection difficult. The optic
disc (OD)-detection algorithm is basically divided into two stages. In the first stage, the
Diagnostics 2023, 1, 0 9 of 18

optic disc is segmented, where candidate regions are calculated using a Fourier transform
followed by morphological operations. If more than one candidate region appears, then
blood vessel segmentation is carried out to calculate the energy of each region. The region
with the maximum energy is marked as the OD [34]. Figure 5 shows the OD coordinates
after applying the algorithm.

Figure 5. Retinal image showing the optical disc coordinates.

The binary thresholded image that we obtain after the abnormal blood vessel extraction
contains both the lesions and the abnormal blood vessels. Neovascularization on disc (NVD)
is graded as the abnormal blood vessels detected at one disc distance (1dd) from the optic
disc coordinates. If D denotes the diameter of the disc then one disc diameter can be
calculated using:
1dd = D + ( D/2) (5)
By applying the one-disc distance, two filter masks are created, one for the NVD case
and the other for the NVE case. These masks are applied to the image for the extraction of
features. Figure 6 shows the filter masks and images to be further processed. True and false
objects in the binary image are first classified on the basis of the 0th moment of the image,
that is, the area of the candidate ROI. Let the characteristic function for the object in the
image be L( x, y). We define:

L( x, y) = 0 f or object and L( x, y) = 1 f or background

so the area of the region can be defined as:


Z Z
A= L( x, y) (6)
Diagnostics 2023, 1, 0 10 of 18

Figure 6. (A,C,E) show the filter mask, binary filtered image, and its map on the original green
channel for NVD, respectively. (B,D,F) show the filter mask, binary filtered image, and its map on
original green channel for NVE, respectively.

The first-order moment, i.e., the center of mass of the objects, is then calculated. Let
the center of mass be denoted by ( x̄, ȳ). Then:
RR
xL( x, y)
x̄ = R R (7)
L( x, y)
RR
yL( x, y)
ȳ = R R (8)
L( x, y)
where x and y are the coordinates of the image. The center of mass is then chosen as the
center of the window for the region of interest. Choosing an appropriate window size is
quite crucial. It is chosen to include all the abnormal blood vessels.

3.5. Feature Selection and Thresholding


It is observed that the abnormal vessels that grow around the optic disc tend to be in
large bunches. They consume more area and are more tortuous. In contrast to that, the
abnormal blood vessels found elsewhere on the retina are comparatively very small and
consume a relatively small area. By considering the properties of these blood vessels the
feature sets used for the classification of NVD are:
• Entropy: Entropy is the measure of uncertainty in a system. Abnormal blood vessels
are fragile and follow no proper pattern. Thus, the regions that contain abnormal
Diagnostics 2023, 1, 0 11 of 18

blood vessels have a high entropy value. If pk is the probability of occurrence of a grey
level k and M is the number of grey levels in the image, then entropy is calculated as:

M
H=− ∑ ( pk ) log2 ( pk ) (9)
k =1

• Energy: Energy is the sum of squares of all pixel intensities within a candidate region
of interest. The energy of the region containing the abnormal blood vessels lies in
between those of the regions that contain normal blood vessels and the bright lesions
or exudates. If g( x, y) is the pixel value in an image then the energy is calculated as:

E= ∑ g(i, j)2 (10)


i,j

• Homogeneity: Homogeneity returns a value that tells the closeness of the distribution
of elements. The homogeneity of abnormal blood vessels lies very close to that
of normal blood vessels but it is away from that of the lesions and exudates. The
abnormal blood vessels originate near normal vessels, while the exudates and bright
lesions can be found anywhere on the retina. The homogeneity is calculated as:

g(i, j)
H= ∑ 1 + |i − j | (11)
i,j

The feature sets used for the classification of NVE are:


• Energy: As a smaller window size was chosen for NVE, it shows a relatively high
energy value in that small area.
• Gradient: The mean gradient magnitude in the candidate region of interest is calcu-
lated by using the Sobel gradient operator. Separate measurements of the gradient
component in each orientation, called Gx and Gy , are calculated. Then the magnitude
of the gradient is given by: q
G ( x, y) = Gx2 + Gy2 (12)

The mean of the gradient magnitude is used as a feature, which is:

1
Mm ag =
nm ∑ G(i, j) (13)
i,j

where m and n are the dimensions of the region of interest.


• Gradient Direction: The directional gradient is the standard deviation of the Sobel
gradient in the candidate region of interest. As the abnormal vessels are much less
defined, are less homogeneous, and have more contrast variation than normal vessels,
this feature is taken in to account. The direction can be calculated as:
Gy
θ ( x, y) = arctan (14)
Gx

and the mean of its standard deviation is:


s
1 (θ − θ̄ )
s= ( ) (15)
m ( n − 1)

The value for the feature sets is chosen inside a fixed window size for each candidate
ROI. The features and characteristics of newly grown vessels lie in between those of the
original vascular pattern and other false detection. Thus, thresholds are applied on both
the upper and lower bounds. After thresholding, there still remains some wrongly detected
Diagnostics 2023, 1, 0 12 of 18

candidate regions of interest, which can be called false-positive detection. A box plot
analysis is used for the further analysis and processing of false-positive detection.

3.6. Post-Processing
The regions that are obtained after ROI processing are grouped into two classes. One
contains the true positive, i.e., the region that actually has neovascularization, and the other
contains the false positive, which does not have neovascularization. A set of features is
applied to both classes. In this analysis, the size of the window for analysis of features is
kept adaptable to the size of the ROI. These features are as below:
1. Mean Intensity ( f 1 ): It is the mean value of pixels within the green plane of the
candidate region.
2. Maximum Intensity ( f 2 ): It is the maximum value of pixels within the green channel
of the candidate region.
3. Mean Skewness ( f 3 ): It is the measure of the lack of symmetry in a candidate region.
It is computed as:
∑ N ( gi − ḡ)3
skewness = i=1 (16)
( n − 1) s3
where g( x, y) is the candidate pixel value, g( x,¯ y) is the mean value pixels, s is the
standard deviation, and N is the number of pixels in the candidate region.
4. Entropy ( f 4 ): It is the value of all pixels in a candidate region and its neighboring
pixels. It is the measure of unpredictability in an ROI.
5. Energy ( f 5 ): It is the sum of the squares of all the pixel values of the green plane
inside a candidate region.
6. Mean Gradient ( f 6 ): It is the mean of the pixels of the edges detected using the Sobel
gradient within the candidate region.
7. Gradient Direction ( f 7 ): It is the standard deviation of the direction of the Sobel
gradient in a candidate region.
8. Mean Intensity of red plane ( f 8 ): It is the mean value of pixels within the red plane of
the candidate region.
9. Mean Intensity of blue plane ( f 9 ): It is the mean value of pixels within the blue plane
of the candidate region.
10. Mean Intensity lightness in LAB color space ( f 9 ): It is the mean value of pixels within
the lightness plane in the LAB color space of the candidate region.
A number of the features mentioned were tested on both sets of candidate ROIs, i.e.,
NVD and NVE. However, not all of them are useful in improving the accuracy of the cases.
Feature selection is very important in any automated system. To select good features for
applying further thresholds, box plots are analyzed. The box plot represents the data in
the form of blocks to show its lowest, highest, and median values [35]. The upper adjacent
limit is found by:
U pperlimit = Q3 + [1.5( Q3 − Q1 )] (17)
and the lower adjacent limit is found by:

Lowerlimit = Q1 − [1.5( Q3 − Q1 )] (18)

where Q1 and Q3 are the first and third quartile, respectively. Figures 7 and 8 show the box
plot for useful NVD and NVE feature sets, respectively. The box plots are analyzed to find
the threshold values. After finding the features and threshold, each candidate region is
subjected to that feature set and the threshold is applied. If the feature value fulfills the
threshold then it is said to be a true detection; otherwise, it is removed.
Diagnostics 2023, 1, 0 13 of 18

Figure 7. Box plots of best features for NVE. (1) True detections, (2) false detections.

Figure 8. Box plots of best features for NVD. (1) True detections, (2) false detections.

3.7. Grading of PDR as NVD or NVE


Once the final thresholds are applied and all abnormal blood vessels are detected,
the system grades the input image as healthy, NVD, NVE, or both based on the location
and distance of these vessels from the OD. If the object detected lies within that one disc
distance, defined in Equation (5), it is marked as NVD. If it lies elsewhere, then it is marked
as NVE. Table 3 shows the grading criteria of PDR.

Table 3. Conditions for grading of PDR.

Grade Condition Class


0 No abnormal blood vessels present Healthy
1 A few abnormal blood vessels present 1dd away from the OD NVE
2 Abnormal blood vessels present within 1dd of OD NVD

4. Results
A dataset is a standard tool for the comparisons and evaluation of different algorithms.
It is very essential for the proper evaluation of medical image-processing-based algorithms.
We evaluated our algorithm on one globally available dataset (MESSIDOR) and one locally
available dataset.
MESSIDOR has been established to facilitate computer-aided diabetic retinopathy
detection [36]. The images in the dataset were acquired with a TopCon TRC NW6 Non-
mydriatic fundus camera with 45o FOV and resolutions of 1440 × 960, 2240 × 1488, and
2304 × 1536 with 8 bits per color plane. A total of 1200 images is contained in this dataset,
which is divided into three subsets of 400 images. Each subset is further divided into
Diagnostics 2023, 1, 0 14 of 18

four parts to facilitate thorough testing. An Excel file accompanies each set that contains
medical findings that are used for testing purposes. These images are graded into different
categories depending on the number, position, and presence of different lesions. Locally,
some data have been collected from the Armed Forces Institute of Ophthalmology (AFIO).
A total of 1200 images from the MESSIDOR database and 20 images from the AFIO database
is used for the evaluation of the proposed algorithm. A detailed description of the database
is given in Table 4.

Table 4. Database description.

Database Images Normal PDR NVD NVE


MESSIDOR 1200 397 37 27 18
AFIO 20 13 7 4 3

In order to perform detailed testing, the algorithm is run on the whole database and
the results are verified with the help of an ophthalmologist. The results are compared with
the ground truth of the abnormal blood vessels marked by ophthalmologists. The results
are also verified with the ground truth attached to the database. A detailed evaluation of
the proposed system is also performed using different statistical evaluation parameters
such as sensitivity, specificity, accuracy, and PPV.

TP
Sensitivity = (19)
( TP + FN )

TN
Speci f icity = (20)
( TN + FP )
TP
PPV = (21)
( TP + FP )
( TP + TN )
Accuracy = (22)
( TP + TN + FP + FN )
where:
• TP are true positives, meaning abnormal blood vessel regions correctly classified as
abnormal.
• TN are true negatives, meaning normal blood vessel regions correctly classified an
normal.
• FP are false positives, meaning normal blood vessel regions wrongly classified as
abnormal.
• FN are false negatives, meaning abnormal blood vessel regions wrongly classified as
normal blood vessel regions.
Table 5 shows a comparison of the proposed system with existing methods for PDR de-
tection.

Table 5. Performance comparison of the proposed system with existing systems for abnormal blood
vessel detection.

Acc/F1
Sr. Method Number of Images Sen Spec
Score
1 Jelinek et al.[13] 27 images – – 0.90
Saranya et
2 50 images from MESSIDOR and DRIVE 0.96 0.89 0.96
al.[15]
3 Lee et al.[17] 137 images from MESSIDOR 0.96 0.99 0.98
Welikala et
4 60 images 0.91 0.92 0.96
al.[18]
5 Garima et al.[15] 799 images 0.95 0.83 0.96
Diagnostics 2023, 1, 0 15 of 18

Table 5. Cont.

Acc/F1
Sr. Method Number of Images Sen Spec
Score
8 Proposed 1200 images from MESIDOR 0.90 1 0.98
– 20 images from AFIO 0.80 1 0.95

Figure 9 illustrates the abnormal blood vessel detection results for the proposed
method. The validity of the proposed system is clearly highlighted. A large number of
images is used for the evaluation of the system. The improvement in results is because
of the accurate extraction of abnormal blood vessels and the optic disc for the detection
of sound feature selection and because of false ROI removal. Table 5 comprehends and
compares the results of various state-of-the-art frameworks with our proposed framework.
On the basis of accuracy, the proposed algorithm yields the highest accuracy on a large
dataset. Other algorithms achieving 0.98 accuracy lack thorough testing on a large dataset.
Moreover, unlike the majority of algorithms cited in the related literature, the proposed
framework not only highlights the diseased cases, but also categorizes them based on
the disease severity, also highlighting the diseased area. Moreover, instead of using deep
learning frameworks to obtain accurate results, our algorithm uses basic image-processing
techniques to reach the final conclusion, which makes it light and less data-hungry, since it
does require pre-training the model, thus giving accurate results even on a small dataset.

Figure 9. Results of grading of PDR: (a) graded as 1 and 2, (b) graded as 1, (c) graded as 1 and 2, (d)
graded as 0.

5. Conclusion
Proliferative diabetic retinopathy (PDR) is an advanced stage of diabetic retinopathy.
In this research, a computerized medical system for the screening of PDR is presented. The
proposed system performs an analysis of retinal images for grading PDR by analyzing box
plots for different sets of features. The proposed system carried out OD detection followed
by region-of-interest detection. The abnormal blood vessel detection stage created a binary
map of candidate regions using filter banks. A detailed feature set based on the properties
of these abnormal blood vessels is created for each candidate region and thresholds are
applied to detect all true abnormal blood vessel regions. A further set of features is then
applied with an adaptive window and the distribution pattern of true and false detection is
Diagnostics 2023, 1, 0 16 of 18

analyzed with box plots. By using the coordinates of the OD and the distance of abnormal
blood vessels from the optic disc, the system graded the input image into three categories,
i.e healthy, NVD, and NVE. The evaluation of the proposed system was performed on
the MESSIDOR and AFIO databases. For evaluation, the statistical measures sensitivity,
specificity, accuracy, and PPV were used. The results showed that the system achieved an
average accuracy of 98.5% and 96.5% for the MESSIDOR and AFIO databases, respectively,
while a PPV and specificity of 99.8% was achieved for both databases. This research’s
contributions are (i) a complete system for the grading of PDR, (ii) improved results by
addressing open issues such as the occurrence of false positives due to the similarity of
abnormal blood vessels to normal blood vessels, (iii) the use of two different techniques
for the extraction of a useful feature set, based on the properties of normal and abnormal
blood vessels, and (iv) the proposal of a method without the need for a classifier, which
saves the time required in training.

Author Contributions: Methodology, N.U.H, M.U.A., N.S.A., and A.A.S.; Validation, J.Z., M.U.A.,
and N.S.A.; Formal analysis, N.U.H.; Writing—original draft, N.U.H. and A.A.S.; Writing—review
and editing, M.U.A. and A.A.S.; Visualization, J.Z.; Supervision, M.U.A. and N.S.A.; Project admin-
istration, M.U.A. and N.S.A.; Funding acquisition, N.S.A. All authors have read and agreed to the
published version of the manuscript.
Funding: Princess Nourah bint Abdulrahman University Researchers Supporting Project number
(PNURSP2023R40), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.
Institutional Review Board Statement:
Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.
Data Availability Statement: The MESSIDOR dataset is available at https://www.adcis.net/en/
third-party/messidor/ (accessed on 29 June, 2023).
Conflicts of Interest: The authors declare no conflicts of interest.

Abbreviations
The following abbreviations are used in this manuscript:

PDR Proliferative Diabetic Retinopathy


DR Diabetic Retinopathy
NVD Neovascularization on Disc
NVE Neovascularization Elsewhere
SVM Support Vector Machine
OD Optic Disc

References
1. Leontidis, G.; Al-Diri, B.; Hunter, A. Diabetic retinopathy: current and future methods for early screening from a retinal
hemodynamic and geometric approach. Expert Rev. Ophthalmol. 2014, 9, 431–442.
2. Arden, G.B.; Ramsey, D.J. Diabetic retinopathy and a novel treatment based on the biophysics of rod photoreceptors and dark
adaptation. In Webvision: The Organization of the Retina and Visual System; Kolb, H., Fernandez, E., Nelson, R., Eds.; University of
Utah Health Sciences Center: Salt Lake City, UT, USA, 2015; Available online: https://www.ncbi.nlm.nih.gov/books/NBK310272/
(accessed on 29 June, 2023).
3. Wong, T.; Cheung, C.; Larsen, M.; Sharma, S.; Simó, R. Diabetic retinopathy. Nat. Rev. Dis. Prim. 2016, 2, 16012.
https://doi.org/10.1038/nrdp.2016.12
4. Shukla, U.V.; Tripathy, K. Diabetic Retinopathy. In StatPearls; StatPearls Publishing: Treasure Island, FL, USA, 2023. Available
online: https://www.ncbi.nlm.nih.gov/books/NBK560805/ (accessed on 22 February 2023).
5. Nentwich, M.M.; Ulbig, M.W. Diabetic retinopathy—Ocular complications of diabetes mellitus. World J. Diabetes 2015, 6, 489–499.
https://doi.org/10.4239/wjd.v6.i3.489.
6. Salamat, N.; Missen, M.M.; Rashid, A. Diabetic retinopathy techniques in retinal images: A review. Artif. Intell. Med. 2019, 97,
168–188.
7. Akram, M.U.; Khalid, S.; Tariq, A.; Javed, M.Y. Detection of Neovascularization in Retinal Images using Multivariate m-Mediods
based Classifier. Comput. Med Imaging Graph. 2013, 37, 346–357.
Diagnostics 2023, 1, 0 17 of 18

8. Leontidis, G.; Al-Diri, B.; Wigdahl, J.; Hunter, A. Evaluation of geometric features as biomarkers of diabetic retinopathy
for characterizing the retinal vascular changes during the progression of diabetes. In Proceedings of the 2015 37th Annual
International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Milan, Italy, 25–29 August 2015; pp.
5255-5259. https://doi.org/10.1109/EMBC.2015.7319577.
9. Mendes, L.; Marques, I.P.; Cunha-Vaz, J. Comparison of Different Metrics for the Identification of Vascular Changes in Diabetic
Retinopathy Using OCTA. Front Neurosci. 2021, 15, 755730. https://doi.org/10.3389/fnins.2021.755730.
10. Das, S.; Kharb, a.K.; Suchetha, M.; Raman, R.; Dhas, E. Deep learning architecture based on segmented fundus image features for
classification of diabetic retinopathy. Biomed. Signal Process. Control 2021, 68, 102600.
11. Goatman, K.A.; Fleming, A.D.; Philip, S.; Williams, G.J.; Olson, J.A.; Sharp, P.F. Detection of new vessels on the optic disc using
retinal photographs. IEEE Trans. Med. Imaging 2011, 30, 972–979.
12. Agurto, C.; Murray, V.; Barriga, E.; Murillo, S.; Pattichis, M.; Davis, H.; Russell, S.; Abramoff, M.; Soliz, P. Multi-scale AM-FM
methods for diabetic retinopathy lesion detection. IEEE Trans Med Imaging 2010, 29, 502–512.
13. Jelinek, H.F.; Cree, M.J.; Leandro, J.J.G.; Soares, J.V.B.; Cesar, R.M., Jr.; Luckie, A. Automated segmentation of retinal blood vessels
and identification of proliferative diabetic retinopathy. J. Opt. Soc. Am. A 2007, 24, 1448–1456.
14. Mudigonda, S.; Oloumi, F.; Katta, K.M.; Rangayyan, R.M. Fractal analysis of neovascularization due to diabetic retinopathy
in retinal fundus images. In Proceedings of the E-Health and Bioengineering Conference (EHB), IEEE, Iasi, Romania, 19–21
November 2015.
15. Saranya, K.B.; Mohideen, S.K. A Novel Approach for the Detection of New Vessels in the Retinal Images for screening Diabetic
Retinopathy. In Proceedings of the 2012 International Conference on Communication and Signal Processing: IEEE Advancing
Technology for Humanity, Chennai, India, 4–5 April 2012. https://doi.org/10.1109/ICCSP.2012.6208394.
16. Roychowdhury, S.; Koozekanani, D.D.; Parhi, K.K. Automated detection of neovascularization for proliferative diabetic retinopa-
thy screening. In Proceedings of the 2016 IEEE 38th Annual International Conference: Engineering in Medicine and Biology
Society (EMBC), IEEE, Orlando, FL, USA, 16–20 August 2016.
17. Lee, J.; Zee, B.C.Y.; Li, Q. Detection of Neovascularization Based on Fractal and Texture Analysis with Interaction Effects in
Diabetic Retinopathy. PLoS ONE 2013, 8, e75699. https://doi.org/10.1371/journal.pone.0075699.
18. Welikalaa, R.A.; Dehmeshki, J.; Hoppe, A.; Tah, V.; Mann, S.; Williamson, T.H.; Barman, S.A. Automated detection of proliferative
diabetic retinopathy using a modified line operator and dual classification. Comput. Methods Programs Biomed. 2014, 114, 247–261.
19. Yu, S.; Xiao, D.; Kanagasingam, Y. Automatic detection of neovascularization on optic disk region with feature extraction and
support vector machine. In Proceedings of the 2016 IEEE 38th Annual International Conference of the Engineering in Medicine
and Biology Society (EMBC), IEEE, Orlando, FL, USA, 16–20 August 2016.
20. Coelho, D.F.G.; Rangaraj, M.R.; Vassil, S.D. Detection of neovascularization near the optic disk due to diabetic retinopathy.
In Proceedings of the 2016 24th European Signal Processing Conference (EUSIPCO), IEEE, Budapest, Hungary, 29 August–2
September 2016.
21. Yu, S.; Xiao, D.; Kanagasingam, Y. Machine learning based automatic neovascularization detection on optic disc region. IEEE J.
Biomed. Health Inform. 2018, 22, 886–894.
22. Argyrios, C.; Hurtut, T.; Cheriet, F. Proliferative diabetic retinopathy characterization based on the spatial organization of vascular
junctions in fundus images. In Proceedings of the 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017),
IEEE, Melbourne, VIC, Australia, 18–21 April 2017.
23. Leeza, M.; Farooq, H. Detection of Severity Level of Diabetic Retinopathy Using Bag of Features Model. IET Comput. Vis. 2019, 13,
523–530.
24. Zhang, L.; Li, Q.; You, J.; Zhang, D. A modified matched filter with double-sided thresholding for screening proliferative diabetic
retinopathy. IEEE Trans. Inf. Technol. Biomed. 2009, 13, 528–534.
25. Qummar, S.; Khan, F.G.; Shah, S.; Khan, A.; Shamshirb, ; S.; Rehman, Z.U.; Khan, I.A.; Jadoon, W. A Deep Learning Ensemble
Approach for Diabetic Retinopathy Detection. IEEE Access 2019, 7, 150530–150539.
26. Elgafi, M.; Sharafeldeen, A.; Elnakib, A.; Elgarayhi, A.; Alghamdi, N.S.; Sallah, M.; El-Baz, A. Detection of Diabetic Retinopathy
Using Extracted 3D Features from OCT Images. Sensors 2022, 22, 7833.
27. Mehboob, A.; Akram, M.U.; Alghamdi, N.S.; Abdul, Salam, A. A Deep Learning Based Approach for Grading of Diabetic
Retinopathy Using Large Fundus Image Dataset. Diagnostics 2022, 12, 3084.
28. Tang, M.C.; Teoh, S.S.; Ibrahim, H.; Embong, Z. Neovascularization detection and localization in fundus images using deep
learning. Sensors 2021, 21, 5327.
29. Tang, M.C.; Teoh, S.S.; Ibrahim, H.; Embong, Z. A deep learning approach for the detection of neovascularization in fundus
images using transfer learning. IEEE Access 2022, 10, 20247–20258
30. Lavanya, S.; Naveen, P. Detection of Retinal Neovascularization Using Optimized Deep Convolutional Neural Networks. J.
Trends Comput. Sci. Smart Technol. 2022, 4, 38–49.
31. Tariq, A.; Akram, M.U. An Automated System for Colored Retinal Image Background and Noise Segmentation. In Proceedings of
the 2010 IEEE Symposium on Industrial Electronics and Applications (ISIEA), Penang, Malaysia, 3–5 October 2010.
32. Antoine, J.P.; Carette, P.; Murenzi, R.; Piette, B. Image analysis with two-dimensional continuous wavelet transform. Signal
Process. 1993, 31, 241–272.
Diagnostics 2023, 1, 0 18 of 18

33. Akram, M.U.; Khan, S.A. Multilayered thresholding-based blood vessel segmentation for screening of diabetic retinopathy. Eng.
Comput. 2011, 29, 165–173. https://doi.org/110.1007/s00366-011-0253-7.
34. Usman, A.; Abbas, S.; Akram, M.U.; Nadeem, Y. A Robust Algorithm for Optic Disc Segmentation from Colored Fundus Images: Image
analysis and Recognition; Springer: Berlin/Heidelberg, Germany, 2014.
35. Nelson, L.S. Evaluating Overlapping Confidence Intervals. J. Qual. Technol. 1989, 21, 140141.
36. MESSIDOR. Available online: https://www.adcis.net/en/third-party/messidor/ (accessed on 29 January 2023).

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.

You might also like