Professional Documents
Culture Documents
J. Swarnalakshmi
Department of Computer Science and Engineering
Hindustan Institute of Technology and Science
Chennai, India
swarnalakshmijr@gmail.com
I. INTRODUCTION
Abstract: Deep learning (DL) enables the creation of
computational models comprising multiple processing
Diabetes is a long-term blood sugar disorder that
layers that learn data representations at multiple levels
of abstraction. In the recent past, the use of deep impairs an individual/s body's capacity to use blood
learning has been proliferating, yielding promising
sugar efficiently. This is primarily due to a decrease
results in applications across a growing number of
fields, most notably in image processing, medical image in insulin synthesis in the body. When the cells stop
analysis, data analysis, and bioinformatics. DL responding to insulin, too much blood sugar is
algorithms have also had a significant positive impact
through yielding improvements in screening, present in the person's bloodstream. This further
recognition, segmentation, prediction, and classification causes health issues such as cardiovascular disease,
applications across different domains of healthcare,
such as those concerning the abdomen, cardiac, loss of eyesight, and renal disease. Hence, the early
pathology, and retina. Given the extensive body of detection of diabetes is crucial as medical
recent scientific contributions in this discipline, a
comprehensive review of deep learning developments in intervention can avert significant repercussions.
the domain of diabetic retinopathy (DR) analysis, viz., Diabetes is classified into Type 1, Type 2 and
screening, segmentation, prediction, classification, and
"Gestational diabetes".
validation, is presented here. A critical analysis of the
relevant reported techniques is carried out, and the Diabetes type 1: Children and teens are most
associated advantages and limitations highlighted,
typically diagnosed with type 1 diabetes. When an
culminating in the identification of research gaps and
future challenges that help to inform the research individual has diabetes type 1, their immune system
community to develop more efficient, robust, and targets the pancreatic cells that produce insulin.
accurate DL models for the various challenges in the
monitoring and diagnosis of DR. Diabetes type 2: Type 2 diabetes is caused when an
individual's body stops reacting to the insulin
keyword: Dataset of retinal images, CNN
produced by the pancreas. The pancreas eventually
quits making enough insulin. It's usually caused by a
mix of hereditary and environmental factors.
fluids. It is important to seek medical help
immediately as if left untreated, it might cause
scarring in the eye. Furthermore, the leaking of these
blood vessels might worsen. Early detection of the
disease can slow down its impact of the disease and
the development of DR into a more advanced
eyesight-threatening form can be prevented.
There are two main stages of DR based on the extent
of damage, these are:
Gestational Diabetes: also known as Diabetes that is this stage of DR, and minute bulges occasionally
caused during pregnancy. This kind of diabetes is erupt from the smaller arteries, which leads to fluid
caused by hormonal changes that occur during and blood being leaked into the retina. The diameter
Placental hormones can reduce the body's sensitivity and change. NPDR can worsen when more blood
to insulin. This can lead to elevated blood sugar vessels are blocked, going from mild to severe. There
d. Inflated Cholesterol with FNN and DNN for image classification. They
1. Suvajit et al., used 1000 retinal pictures which they use transfer learning, which is used to retrain the
were subjected to the back propagation technique in rearmost layer of a deep neural network. The Shallow
neural networks, deep neural networks, and neural network method has an accuracy of 72.35%,
67.05%, and 69.03% on the train, validation, and test
sets respectively. The transferred learning method has
an accuracy of 80.85%, 80.60%, 77.87% for the train
set, validation set and test set respectively. According
to the accuracies that were obtained during the
experimentation we can observe that the second
method gave a better accuracy even for the less Developing a Deep Learning model for Diabetic
frequent classes. Retinopathy involves several steps, from data
collection to model evaluation. Here’s a general
workflow:
3. Khalid et al, collected data from 327 patients using Data Collection:
a random sampling technique with proportional Gather a dataset of retinal images labeled with
allocation. They have used a 10-fold cross-validation diabetic retinopathy stages. Myelinated nerve fiber,
resampling technique that was repeated 100 times. In Preretinal hemorrhage, Normal, VKH disease. This
the end, the accuracy of the respective classifier from dataset should be sufficiently large and diverse to
each iteration is averaged for that specific run. DR capture various manifestations of the disease. dataset
was classified using different methods such as SVM, containing retinal images labeled with diabetic
LDA, KNN, RF and RRF. RRF outperformed the retinopathy stages, as well as other conditions like
other classifiers by achieving an accuracy of 86%. myelinated nerve fiber, Preretinal hemorrhage, and
The algorithm also identified HbA1c and duration of VKH disease, one may embark on a quest through the
diabetes as the most discriminating interpretable vast realm of online data repositories, with Kaggle
features for classifying DR. being a prominent kingdom to explore. Within the
vast corridors of Kaggle's repositories lie datasets that
III. WORK FLOW: may hold the visual tapestry of retinal images
adorned with the intricate patterns of diabetic
retinopathy and its varying stages, alongside the
peculiar manifestations of myelinated nerve fiber,
Preretinal hemorrhage, and VKH disease.
Data Preprocessing:
Resizing images ensures uniformity in Grayscale conversion is relevant when working with
dimensionality, akin to sculpting raw marble into color images. In some cases, color information may
uniform blocks. This process not only conserves not be necessary for the task at hand, or using
computational resources but also facilitates seamless grayscale images can reduce computational
integration into machine learning frameworks. complexity.