Semester: IV/VI/VIII Review: 4.1/ 6.1/ 8.
Title of the Project: Retinal Blindness Detection
Domain: Mentor Name: Dr. Ravita Mishra
Group Members:
Member 1: Arnav Yadav
Member 2: Arjun Singh Silwal
Member 3: Prathamesh Shetty
Member 4: Tarunkumar Sharma
Content
● Overview
● Introduction
● Problem Statement
● Objectives
● Requirements
● Literature Survey
● Proposed System
● Proposed Design (along with UML Diagrams)
● Implementation
● Results and Analysis
● Conclusion
● References
Introduction to Project
Diabetes and aging pose major threats to vision worldwide through two prevalent
r Rates
conditions: Diabetic Retinopathy (DR) and cataracts. DR develops when uncontrolled blood
sugar damages retinal blood vessels, while cataracts—affecting over 65 million people
globally—occur when the eye's lens becomes cloudy with age.
Both conditions progress silently, with patients remaining unaware until significant vision
loss occurs. This asymptomatic progression, combined with limited access to
ophthalmologists in many regions, creates an urgent need for automated screening tools.
Our project addresses this challenge with a neural network that detects both conditions
from fundus images. By combining CNN with Autoencoder technology, we've developed a
model that achieves high accuracy in early detection. This cost-effective approach expands
access to screening and could substantially reduce preventable blindness worldwide.
Problem Statement
Diabetes mellitus and aging are major causes of vision impairment, leading to diabetic
r Rates(DR) and cataracts, respectively. While DR results from vascular damage,
retinopathy
cataracts form due to protein changes in the lens. Both progress silently until significant
vision loss occurs.
Early detection is crucial, but traditional diagnostics require specialized equipment and
experts, limiting access in underserved areas. Our project tackles this by developing CNN
architectures enhanced with Autoencoders to detect both DR and cataracts from fundus
images. This dual-diagnosis approach improves screening efficiency and accessibility,
helping to prevent avoidable blindness worldwide.
Objectives of the project
To develop a machine learning-based approach for early detection of Diabetic Retinopathy (DR) and Cataracts
● Detect DR and cataracts from fundus images using Convolutional Neural Networks (CNNs).
● Identify different stages of DR (No DR, Mild, Moderate, Severe, Proliferative DR) and detect cataract presence.
To reduce the dependency on manual diagnosis and speed up the screening process
● Traditional detection of DR and cataracts requires ophthalmologists and specialized equipment.
● Our system aims to provide an automated, fast, and scalable solution for both conditions.
To enhance the accuracy of DR and cataract classification using advanced Deep Learning techniques
● Use CNN architectures to extract key features such as blood vessel abnormalities, lesions, and lens opacity.
● Optimize the model using data augmentation, feature selection, and hyperparameter tuning for improved
detection.
Requirements of the system (Hardware,
software)
Hardware Requirements
1️. Processor (CPU/GPU)
● Minimum: Intel Core i5 (8th Gen) or AMD Ryzen 5
● GPU: NVIDIA GTX 1050 (Minimum) | NVIDIA RTX 3060/4090 (Recommended) for faster deep
learning computations.
2️. RAM (Memory)
● Minimum: 8GB RAM
3️. Storage (SSD/HDD)
● Minimum: 256GB SSD or 500GB HDD
Statement
Requirements of the system (Hardware,
software)
Software Requirements
1. Operating System : Windows 10/11 (64-bit)
2. Programming Language : Python 3.x (Preferred: Python 3.7 or later)
3. Development Environment (IDE) : Google Colab for testing and model training and VS Code
for writing and debugging Python scripts.
4. Database : MySQL / Firebase (if storing patient records and predictions).
5. Web Framework : Flask for creating a web-based application for DR detection
Literature Survey
Project Author & Detail Advantages Drawbacks Summary
Title
Understanding S. Albawi, T. A. Mohammed Comprehensive Limited This paper provides a
of a and S. Al-Zawi, Explanation: The paper Implementation foundational overview of
convolutional provides a clear, step-by- Details: While the paper Convolutional Neural
lem Statement
neural network "Understanding of a
convolutional neural
step breakdown of CNN
architecture, making
complex concepts
explains CNN concepts
thoroughly, it lacks
specific code examples
Networks (CNNs),
systematically explaining
network," 2017 their architecture and
Problem Statement
International Conference
on Engineering and
accessible to newcomers
in the field.
or implementation
guidelines for different
mathematical principles. It
covers essential
Technology (ICET), Antalya, frameworks. components —convolutional
Mathematical
Turkey, 2017, pp. 1-6, layers, pooling layers, fully
Foundation: Presents Outdated
doi:10.1109/ connected layers—along
the mathematical basis Architectures: Being
ICEngTechnol.2017.830818 for CNNs in an published in 2017, the with activation functions
6. approachable manner, paper doesn't cover and backpropagation.
bridging the gap between more recent CNN Designed as an educational
theory and practical architectural innovations resource, it uses clear
implementation. and improvements visuals and accessible
developed in subsequent explanations to help
years. newcomers understand this
transformative deep
learning architecture for
Literature Survey
Project Author & Detail Advantages Drawbacks Summary
Title
Diabetic Md. Nahiduzzaman, Md. High Accuracy: The Limited Dataset The authors employ Contrast
Robiul Islam, Md. Omaer framework Diversity: The study Limited Adaptive Histogram
retinopathy
Faruq Goni, Md. Shamim demonstrated primarily focuses on Equalization (CLAHE) to
identificatio
lem Statement Anower, Mominul Ahsan, superior accuracy in two datasets, which enhance image quality and
n using Julfikar Haider, Marcin classifying DR levels,
may not capture the
Convolutional Neural Network
parallel full variability of DR
Kowalski, achieving 97.27% on (CNN) for feature extraction.
Problem Statement
convolutiona
l neural Expert Systems with
the APTOS 2019
dataset.
manifestations across
different populations. These features are then
classified using an Extreme
network Applications, Potential Learning Machine (ELM). The
based Volume 217, Efficiency: Utilizing a Overfitting: High proposed framework was
2023, lightweight parallel accuracy on the given evaluated on two datasets:
feature datasets raises the
119557, CNN and ELM reduces the Kaggle DR 2015
extractor ISSN 0957-4174, the number of question of potential competition dataset (Dataset-
and ELM https://doi.org/10.1016/ parameters and overfitting; further 1) with 34,984 images and
classifier validation on external
j.eswa.2023.119557. layers, leading to the APTOS 2019 dataset with
datasets is necessary
(https://www.sciencedirect. shorter processing to confirm 3,662 images.
com/science/article/pii/S09 times.
57417423000581) generalizability.
Literature Survey
Project Author & Detail Advantages Drawbacks Summary
Title
BiRA-Net: Improved Complexity: BiRA- This research paper proposes a
Z. Zhao et al., "BiRA-Net: classification Net is a complex deep learning architecture called
Bilinear BiRA-Net for diabetic retinopathy
Bilinear Attention Net for accuracy: BiRA-Net architecture that
lem Statement
Attention Diabetic Retinopathy achieves a higher requires a large (DR) grading. The architecture
Net for Grading," 2019 IEEE combines a ResNet for feature
average classification amount of
extraction, an attention
Diabetic International Conference accuracy compared to computational
Problem Statement
Retinopathy
on Image Processing (ICIP),
Taipei, Taiwan, 2019, pp.
state-of-the-art
methods.
resources and data
to train.
mechanism to focus on key
features, and a bilinear model for
fine-grained classification. A new
Grading 1385-1389, doi: loss function, called grading loss,
10.1109/ICIP.2019.880307 Fine-grained Dependence on is also proposed to improve
4. keywords: classification: The high-quality data: training convergence. The paper
{Diabetes;Feature bilinear model used in BiRA-Net requires evaluates the performance of
extraction;Retina;Retinopat BiRA-Net enables fine- high-quality data to BiRA-Net on a dataset from
hy;Training} grained classification, train, which may be Kaggle and compares it with
https://ieeexplore.ieee.org/ state-of-the-art methods,
which is essential for challenging to
stamp/stamp.jsp?tp=&arnu showing that BiRA-Net
DR grading. obtain, especially in outperforms them in terms of
mber=8803074&isnumber
=8799366 low-resource average classification accuracy,
settings. macro F1, and micro F1.
Literature Survey
Project Author & Detail Advantages Drawbacks Summary
Title
Detection of S.B. Saleema Parvin. A. High accuracy: The Limited dataset This study presents a CNN-
CNN-AE model diversity: Only Autoencoder architecture for
Cataract Devendran (2024). cataract detection in fundus
achieved 97.38% uses one dataset
lem Statement
Disease testing accuracy for from Chinese images. Using Autoencoder as
Using Detection of Cataract transfer learning and Inception-
right eye images and hospitals/medical
Disease Using V2, the model achieved 97.38%
Convolution 92.84% for left eye centers
Problem
Neural
NetworkStatement
Convolution Neural
Network with
with
images
Lack of
accuracy on right eye images
from a 6,392-image dataset. The
approach enhances feature
Autoencoder. Enhanced feature comparative extraction while reducing
Autoencoder International Journal of learning: evaluation: computation. Despite promising
Intelligent Systems and Successfully leverages Limited comparison results, the performance gap
Applications in the autoencoder to with other state-of- between right and left eyes and
extract more relevant the-art methods limited dataset diversity suggest
Engineering, 12(21s), the need for additional validation
2878 features from fundus
across different populations.
images
–. Retrieved from
https://ijisae.org/index.p
hp/IJISAE/article/view/59
16
Proposed System
The proposed system for our idea is as follow:
Input Module:
● The user uploads a retinal fundus image through the application.
Preprocessing Module:
● The uploaded image undergoes enhancement, noise removal, and resizing to ensure consistency.
Prediction Module:
● The processed image is passed through the trained CNN model for classification.
● The model predicts whether the patient has DR (and its severity) and/or Cataract.
Output Module:
● The system generates a detailed diagnosis report with classification results and confidence scores.
● If DR or Cataracts are detected, the system suggests further medical consultation.
Deployment & Accessibility:
● The system will be accessible via a web app for real-time analysis.
Proposed Design
Proposed Design as per current understanding of Machine Learning is as follows:
Data Collection & Preprocessing:
● Fundus images of the retina will be collected from publicly available datasets and medical sources.
● Images will be enhanced using contrast adjustments, noise reduction, and normalization techniques.
● Data augmentation (rotation, flipping, zooming) will be applied to improve model generalization.
Deep Learning Model Development:
● A Convolutional Neural Network (CNN) will be used as the primary model architecture.
● The model will be trained to classify images into different DR stages and cataract presence or absence.
Model Optimization & Evaluation:
● Hyperparameter tuning will be performed to optimize learning rate, batch size, and architecture depth.
● The model will be evaluated using metrics like accuracy, precision, recall, and F1-score.
● Cross-validation will be used to ensure robustness and reduce overfitting.
Proposed Design
Implementation
Result Analysis
Conclusion
References
● S. Albawi, T. A. Mohammed and S. Al-Zawi, "Understanding of a convolutional neural network,"
2017 International Conference on Engineering and Technology (ICET), Antalya, Turkey, 2017, pp.
1-6, doi:10.1109/ICEngTechnol.2017.8308186.
● Md. Nahiduzzaman, Md. Robiul Islam, Md. Omaer Faruq Goni, Md. Shamim Anower, Mominul
Ahsan, Julfikar Haider, Marcin Kowalski, Expert Systems with Applications,Volume
217,2023,119557, ISSN 0957-4174, https://doi.org/10.1016/j.eswa.2023.119557.
● Z. Zhao et al., "BiRA-Net: Bilinear Attention Net for Diabetic Retinopathy Grading," 2019 IEEE
International Conference on Image Processing (ICIP), Taipei, Taiwan, 2019, pp. 1385-1389, doi:
10.1109/ICIP.2019.8803074.
● S.B. Saleema Parvin. A. Devendran (2024). Detection of Cataract Disease Using
Convolution Neural Network with Autoencoder. International Journal of Intelligent Systems
and Applications in Engineering, 12(21s), 2878 –. Retrieved from
https://ijisae.org/index.php/IJISAE/article/view/5916
Publications (if any)