PUDUCHERRY TECHNOLOGICAL UNIVERSITY

Department of Computer Science and Engineering

HYBRID DEEP LEARNING MODEL FOR

DIABETIC RETINOPATHY DETECTION
CS233- Project Work
First Review

Guided by, Presented by,

Dr. N. Sivakumar Madhumidaa R – 20CS1069
Professor Tamiri Mounya Krishna – 20CS1113
Dept Of CSE, PTU. Vijitha Mittapalli – 20CS1123
AGENDA

• Zeroth review(justification)
• Introduction
• Problem Definition
• Objective
• Literature Survey
• Existing Work
• Overall Limitation
• Proposed Work
• Dataset
• Input/Output of the system
• Metrics
• Hardware/Software Requirements
• Conclusion
• Plan of action
• References
ZEROTH REVIEW COMMENTS(JUSTIFICATION)

• Feature Used to detect the severity(in Manual grading)

- It depends on the swelling range and the cotton wool spots in the retina.
Nerve size doesn’t play role in case of detection.

• Changing the title in the name of Classification

- “Hybrid Deep Learning Model for Diabetic Retinopathy Detection”
INTRODUCTION

• No. of people affected by Diabetes has increased rapidly over the

past two decades.

• Diabetic Retinopathy(DR) is the condition of the people who have

diabetes for longer period, which poses significant threat to vision.

• Early diagnosis and accurate detection of DR are vital to prevent

loss of vision.
PROBLEM DEFINITION

• At present, Manual examining of fundus images used to detect the

severity of DR.

• Traditional manual grading is time consuming and resource intensive.

• Identifying DR in its early stages can be challenging as the symptoms may

not be readily apparent.

• Microscopic damage to tiny retinal blood vessels, causing fluid leakage

and distinctive features form the core of the diagnostic challenge.
OBJECTIVE

• To detect and classify DR images in

its early stages using deep learning
and transfer learning models.

• To improve the
performance(enhance the speed and
accuracy) by utilizing Artificial
Intelligence(AI), Deep
Learning(DL) models which will Stages of DR
find features in retinal images.
LITERATURE SURVEY

S.NO PAPER TITLE TECHNIQUE DATA SET LIMITATIONS ADVANTAGES

1. Pratt, Harry, et CNN Eye PACs Low sensitivity High specificity
al."Convolutional neural
networks for diabetic
retinopathy."
Procediacomputer science
90 (2016): 200-205. Elsevier

2. Supriya, Seema, Zia. Two CNN APTOS (Asia Complexity in Parameter

“Diabetic Retinopathy architectures: Pacific Tele- Implementation Efficiency
Detection using Deep VGG16 and Ophthalmology
Learning”. 2020 IEEE. DenseNet121. Society) 2019
Blindness Detection.
3. Bhimavarapu, U.; Chintalapudi, ResNet-50 model Two fundus image Localization Efficient Accuracy
N.; Battineni, G. Automatic
Detection and Classification of datasets: APTOs and Results
Diabetic Retinopathy Using the Kaggle
Improved Pooling Function in the
Convolution Neural
Network. Diagnostics 2023, 13,
2606.
S.NO PAPER TITLE TECHNIQUE DATASET LIMITATIONS ADVANTAGES
.
4. “Early detection of diabetic RestNet -34 Ultrawide field Dependent on Ease in training
retinopathy based on deep learning architecture fundus single data source for huge retinal
and ultra-wide-field fundus photographs from for images surface image
images”.
Kangrok Oh,Hae Min St. Mary’s hospital dataset
Kang,Dawoon Leem, Hyungyu in South Korea
Lee, Kyoung Yul Seo and
Sangchul. Jan 2021 PMC

5. “Diabetic retinopathy detection CNN Architecture Eye PACs Overfitting Performance

using Deep convolutional concerns enhancement
neural networks”. Darshit,
Aniket, Deep Sidhpura. 2016
IEEE
EXISTING WORK
AN ENSEMBLE DEEP LEARNING TECHNIQUE FOR DETECTION AND CLASSIFICATION OF DIABETIC
RETINOPATHY

• This Paper proposes a deep learning ensemble model for the detection and classification
of diabetic retinopathy. Input data is the retinal images used are fundus images that are
acquired under heterogeneous imaging conditions.

Ensembled Architecture
EXISTING WORK

• MODELS USED:
ResNeXt: Utilizes a "split transform merge" technique to improve feature
representation.
DenseNet: Incorporates dense connections for effective information transfer between
layers.
It combines ResNeXt and modified DenseNet for enhanced generalization.

ResNext and Densenet Architectures

EXISTING WORK

Module 1: Data preprocessing

• This module loads and preprocesses the data and splits the dataset into
training, validation and test sets. It involves resizing of the datasets and
augmentation

Module2: Model Training

• This module trains the individual base models and implementing of each
base model separately in order to evaluate the performance
EXISTING WORK

Module3: Ensemble construction and Evaluation

• This module combines the predictions of individual base models to create

an ensemble and Evaluates the performance of the ensemble model on the
validation set
DATASET

• Utilized the publicly accessible fundus images from the Kaggle dataset, supplied by the
Asia Pacific Tele-Ophthalmology Society (APTOS) for the 2019 Blindness Detection
Competition.
• The dataset consists of 5584 high-resolution images. These images were evaluated for
diabetic retinopathy severity on a scale of 0 to 4 by clinicians, with severity levels
spanning from "No DR" to "Proliferative DR."
EXISTING WORK

MODULE1: DATAPREPROCESSING IMPLEMENTATION

EXISTING WORK

MODULE2: TRAINING MODELS IMPLEMENTATION

• INPUT OF THE SYSTEM
Raw fundus images
• OUTPUT OF THE SYSTEM
DR severity level
DENSENET MODEL RESULT ANALYSIS
RESNEXT MODEL RESULT ANALYSIS
PROPOSED WORK

• We use a hybrid deep learning approach using

deep convolutional neural network (CNN)
method and two VGG(Visual Geometry Group)
network models (VGG16 and VGG19) for
diabetic retinopathy detection and classification.

• The images are pre-processed by performing

techniques like resizing, augmentation to
enhance the features of the image.

• The hybrid model is trained, followed by

hyperparameter tuning.
PROPOSED WORK

Hybrid model Architecture

MODULES

• Data Preprocessing
• Model Training
• Performance evaluation and Hyperparameter Tuning
• Deployment and Integration
DATA SET

•The publicly available fundus images from the Kaggle

dataset, provided by Asia Pacific Tele-Ophthalmology
Society (APTOS) for the 2019 Blindness Detection
Competition were used.

•The dataset comprises of 5584 high-resolution images,

rated for diabetic retinopathy severity on a scale of 0 to 4
by clinicians.

• Severity levels range from "No DR" to "Proliferative

DR."
INPUT/OUTPUT OF THE SYSTEM

Preprocessing stage:
Input : Raw fundus images from the APTOS2019 test dataset, which
includes images from various clinical settings and different individuals.
Output: Standardized and augmented fundus images (128x128 pixels)
with reduced noise and improved features
Feature extraction and classification stage:
Input : Preprocessed fundus images of size 128x128 pixels
Output: Predicted severity labels for diabetic retinopathy, ranging from
0 to 4
METRICS

• The proposed hybrid approach's classification performance is evaluated using key

metrics, including accuracy, precision, sensitivity (recall), specificity, and F1 score.
• Additionally, a confusion matrix is employed to visualize the performance across
multiple classes in the context of diabetic retinopathy classification.
• Specificity = TP∕(TP + FP).
Sensitivity (recall) = TP∕(TP + FN).
Precision = TP∕(TP + FP).
Accuracy = (TP + TN)/(TP + TN + FP + FN) - correct classification proportion.
F− score = 2 × ( precision × recall )/ (precision + recall).
HARDWARE & SOFTWARE REQUIREMENTS

Hardware Requirements:
• RAM : Minimum 4GB or above
• Hard disk or SSD 120GB
• I5 Processor 5th generation or above and GPU​

Software Requirements:
• Cloud Tool: Google Colab (cloud based jupyter notebook)
• Libraries : TensorFlow, Keras , OpenCV, Pandas, Numpy , Matplotlib
CONCLUSION

• The Proposed work strives to advance Diabetic Retinopathy (DR) cancer

CT image classification by integrating cutting-edge techniques. It
introduces a hybrid deep learning approach using (CNN, VGG16, VGG19)
and multi-class classification for early-stage diabetic retinopathy
detection. It significantly boosts the accuracy of diabetic retinopathy CT
image classification and emphasis on cost-effectiveness.
PLAN OF ACTION

>>Data >>Ensemble model

preprocessing of >>Training of Hybrid
construction
>>Requirement model and result
>>Project Existing work
analysis analysis
implementation >>Data
>>Planning
>>Training of preprocessing of
>>Report
individual models proposed model
REFERENCES

1. Anila Sebastian, Omar Elharrouss, Somaya Al-Maadeed and Noor Almaadeed “A Survey on
Deep-Learning-Based Diabetic Retinopathy Classification”, Jan, 2023.
2. Khan, Sharzil Haris, Zeeshan Abbas, and SM Danish Rizvi. "Classification of diabetic retinopathy
images based on customised CNN architecture." 2019 Amity International Conference on
Artificial Intelligence (AICAI). IEEE, 2019.
3. Wang, Xiaoliang, et al. "Diabetic retinopathy stage classification using convolutional neural
networks." 2018 IEEE International Conference on Information Reuse and Integration (IRI).
IEEE, 2018.
4. AbdelMaksoud, E., Barakat, S. & Elmogy, M. A computer-aided diagnosis system for detecting
various diabetic retinopathy grades based on a hybrid deep learning technique. Med Biol Eng
Comput 60, 2015–2038 (2022).
5. Butt, M.M.; Iskandar, D.N.F.A.; Abdelhamid, S.E.; Latif, G.; Alghazo, R. Diabetic Retinopathy
Detection from Fundus Images of the Eye Using Hybrid Deep Learning
Features. Diagnostics 2022, 12, 1607. https://doi.org/10.3390/diagnostics12071607
THANK
YOU!