1.

IDENTIFICATION AND CLASSIFICATION OF MICROANEURYSMS FOR EARLY

DETECTION OF DIABETIC RETINOPATHY

M.Usman Akrama, Shehzad Khalidb, Shoab A.Khana

In this paper, we propose a three-stage system for early detection of microaneurysms (MAs)
using filter banks. In the first stage, the system extracts all possible candidate regions for MAs
present in retinal image. In order to classify a candidate region as MA or non-MA, the system
formulates a feature vector for each region depending upon certain properties, i.e. shape, color,
intensity and statistics. We present a hybrid classifier which combines the Gaussian mixture
model (GMM), support vector machine (SVM) and an extension of multimodel medoid based
modeling approach in an ensemble to improve the accuracy of classification.

The common symptoms of diabetic retinopathy are blurred vision, floaters and flashes, and
sudden loss of vision.

A healthy retina contains blood vessels, optic disc and macula.

DR is broadly divided into two stages, i.e. non-proliferative DR (NPDR) and proliferative DR
(PDR

NPDR, also known as background DR, occurs when diabetes damages the blood vessels
inside the retina causing leakage of blood and fluid on the surface of retina. The functionality of
retina reduces as it becomes wet and swollen due to this leakage. NPDR may contain different
signs of retinopathy such as microaneurysms (MAs), hemorrhages (H), hard exudates (HE) and
soft exudates or cotton wool spots (CWS).
Depending upon the presence and quantity of these lesions, NPDR is further divided into three
stages, i.e. mild, moderate and severe.

MAs are the first sign of NPDR and are caused by the focal dilatations of thin blood vessels.
MAs are of small size, almost round in shape and red in color. The next sign of DR is H which is
also referred to as dot or blot H. When the wall of thin vessels or MAs is sufficiently weakened,
it may rupture and give rise to an H.

Overall, we divide the proposed system into three classical stages, i.e. candidate region
extraction, feature vector formation and classification. The candidate region extraction phase
enhances the regions containing red lesions and extracts all possible candidate regions for MAs.
It also removes all possible blood vessel pixels to minimize spurious regions. In stage 2, feature
vectors are calculated for each extracted region using some distinct properties. Finally, the
classifier separates out the MA and non-MA regions.

The proposed algorithm for the candidate region extraction works in three phases. In phase 1, it
improves the contrast of dark regions by using mathematical morphological operations, contrast
normalization and filter banks. Phase 2 performs blood vessel enhancement and segmentation.
The last phase of candidate region extraction eliminates all blood vessel pixels from candidate
pixels in order to reduce the spurious regions and to increase the accuracy of classifier.

2. A DEEP LEARNING ENSEMBLE APPROACH FOR DIABETIC RETINOPATHY

DETECTION

SEHRISH QUMMAR, FIAZ GUL KHAN, SAJID SHAH, AHMAD KHAN

In this research, we used the publicly available Kaggle dataset of retina images to train an
ensemble of five deep Convolution Neural Network (CNN) models (Resnet50, Inceptionv3,
Xception, Dense121, Dense169) to encode the rich features and improve the classification for
different stages of DR. The experimental results show that the proposed model detects all the
stages of DR unlike the current methods and performs better compared to state-of-the-art
methods on the same Kaggle dataset.

The different preprocessing steps that we perform on input dataset before giving it to the model
are shown in Figure 2 We use the Kaggle dataset which contains 35126 color fundus images;
each is of size 3888 × 2951. It contains the images from five different classes based on the
severity of diabetic retinopathy (DR). Dataset shows the distribution of sample images in
different classes of the Kaggle dataset. The distribution of different classes is shown in the first
row of Table 1, which is perfectly imbalanced. Training of deep networks with imbalance data
leads to classification biasness. In the first preprocessing step, we resize each input image to
786 × 512 by maintaining the aspect ratio to reduce the training overhead of deep networks.
Moreover, for balancing the dataset we performed up-sampling and down-sampling. The up-
sampling is performed with augmentation of minority classes by randomly cropping patches, of
size 512 × 512, followed by flipping and 90o rotation to balance the samples of different classes,
enrich dataset and avoid. In down-sampling extra instances of majority classes are removed to
meet the cardinality of the smallest class. In the resultant distributions, before flipping and
rotation, each image is mean normalized to avoid features biasness and speed-up training time.
The dataset is divided into three parts: training, testing, and validation sets with ratio 64% and
20% and 16% respectively. During training, the validation set is used to check and reduce the
over-fitting.

Ensemble method is a meta-algorithm that combines several machine learning techniques into
one predictive model.

Require: Fundus Images (X, Y ); where Y = {y/y ∈ {Normal, Mild, Moderate, Severe, PDR}}

Output: The trained model that classifies the fundus image x ∈ X 1

Perform Preprocessing:

 Resize the image to dimension 786 × 512

 Perform augmentation: Randomly crop five patches, of size 512 × 512, of each image and
perform flip flop and 90 deg. rotation
 Mean normalize the each image
It can be used for different objectives such as to decrease variance (Bagging), bias (boosting), or
improve predictions (stacking). Stacking is a model used to combine information from multiple
predictive models to generate a new model. The stacked approach often outperforms individual
models due to its soothing nature. Stacking highlights each base model where it performs best
and discredits each base model where it performs poorly. For this reason, stacking is most
effective when the base models are significantly different. For this reason, we used stacking to
improve the prediction of our model, which is evident from our results.
3. DIAGNOSIS OF DIABETIC RETINOPATHY USING MACHINE LEARNING
TECHNIQUES
Mehmet Akif ÇİFÇİ
In this study, machine learning (ML) techniques are used to diagnose DR early. These are PNN,
SVM, Bayesian Classification, and K-Means Clustering. These techniques will be evaluated and
compared with each other to choose the best methodology. A total of 300 fundus photographs are
processed for training and testing. The features are extracted from these raw images using image
processing techniques. After an experiment, it is concluded that:

 PNN has an accuracy of about 89%

 Bayes Classifications 94%
 SVM 97%
 K-Means Clustering 87%
The preliminary results prove that SVM is the best technique for early detection of DR.
4. DIAGNOSIS OF DIABETIC RETINOPATHY USING DEEP NEURAL NETWORKS
To automate the diagnosis of DR and provide appropriate suggestions to DR patients, we have
built a dataset of DR fundus images that have been labeled by the proper treatment method that
is required. Using this dataset, we trained deep convolutional neural network models to grade the
severities of DR fundus images. We were able to achieve an accuracy of 88.72% for a four-
degree classification task in the experiments. We deployed our models on a cloud computing
platform and provided pilot DR diagnostic services for several hospitals; in the clinical
evaluation, the system achieved a consistency rate of 91.8% with ophthalmologists,
demonstrating the effectiveness of our work.
Preprocessing
We need to process the acquired fundus images from different sources and change them into a
uniform format through the following steps:

1) Size Normalization
The first step is to resize different images into a uniform scale so that all fundus areas in different
images have the same diameter. The black borders on each side of the fundus image are removed
at the outset by summing the images horizontally and vertically and discarding regions that
correspond to values under a selected threshold. Then, the images are resized to fixed
dimensions.

2) Shape Normalization
Some fundus images are complete circles, whereas others may lack the top and bottom margins.
In addition, many devices capture a small notch on the edge of the circle. To unify the shapes of
these images, we use a mask that contains the largest common area of all images from different
sources to obscure unwanted parts of the image.

3) Color Normalization
After the shape of each image is normalized, its color must be tuned because different devices
may produce images with different color temperatures, and the illumination conditions can vary.
Our method of color tuning is simple: we shift each of the RGB channels of a fundus image to a
pre-calculated mean and truncate the values above 255.

Augmentation
Having sufficient training data is the key for training a neural network successfully;
unfortunately, this requirement is seldom satisfied in most neural network applications. For
medical imaging applications, the lack of data is more significant because of the cost of the
annotations, and because of the imbalance in the occurrence between diseases. To mitigate
shortages in data and fully utilize the data that are available, certain data augmentation
techniques must be carried out; in our experiment, we used the Augmentator software package.
Specifically, we augmented our data through the following means:

 Flip the image horizontally

 Flip the image vertically

 Randomly rotate the image in the range of [−25,25] degrees

 Randomly zoom in or out in the range of [0.85,1.15]

 Randomly distort the image

All of these methods were combined for augmenting each image, and a probability of 0.5 was
used to determine whether or not to perform each of them.
5. DETECTING CLINICAL FEATURES OF DIABETIC RETINOPATHY USING
IMAGE PROCESSING
Nimmy Thomas, Therese Yamuna Mahesh
The paper deals with detecting the diabetic retinopathy retinal changes. The retinal images are
first subjected to pre-processing techniques like colour normalization and enhancement process.
There may exist different kinds of lesions caused by diabetic retinopathy in a diabetic patient’s
eye such as micro aneurysm, hard exudates, soft exudates, hemorrhage etc. Automated analysis
of the fundus (retinal image) image is very much essential and will be of help to facilitate the
clinical diagnosis.
The proposed system includes the feature selection of the retinal image by detecting the
exudates (soft and hard), blood vessels, micro aneurysms, optic disc, and hemorrhages