Professional Documents
Culture Documents
Guided By
Dr. Chetan Gondaliya
Presented by
Amin Yash Dharmeshkumar (21084341001)
M.Sc.IT. (AI&ML)
Project Profile
Existing systems for eye disease classification use deep learning techniques, such as
CNNs and transfer learning.
Large datasets of labeled retinal images have been used to train and validate these
systems.
Existing systems have shown promising results in detecting and classifying various eye
diseases.
Need For New System
The data we are using, was also used by some researchers previously but they only
used 2-3 disease for classification.
Some of them also used the whole data which include all diseases but the accuracy was
comparatively low.
So, our main motive is to get good accuracy on all diseases, for that we will apply some
filters that will help the model to detect the disease easily.
Function Specification
Client: -
User Registration and Login
Uploading Images
Display Classification Results
Admin: -
Admin Login
User Management:
Prediction management
Client Configuration Server Configuration
Hardware :- Hardware :-
PC or Laptop i5 processor minimum
8gb RAM
Software :-
Web browser like Mozilla
Firefox, Google chrome, etc. Software :-
Python 3.9
Web browser
Google Colab
Literature Survey
K.M.A Salam et al. :- In this paper, author used different cnn architecture like resnet
34, efficient net, MobilenetV2 and VGG16.The VGG-16 model provided the best
accuracy of 97.23% in classifying ocular diseases from fundus photographs, while the
other models also performed well. The proposed method can generate more impressive
results on lower computational power.[1]
Nadia Kanwal et al. :- In this paper, author used simple machine learning approach to
classify the ocular disease from the tabular data which was symptoms recorded in the
form of standard taxonomies directly by the medical experts. They used multiple
classification algorithms but Tree-based methods performed better than the artificial
neural network. [2]
Md Shakib Khan et al. :- In this paper, author used VGG-19 model was used to
classify ocular diseases, achieving high accuracy rates for normal versus myopia,
cataract, and glaucoma. The model could be used to create a consumer-friendly ocular
disease categorization system and could benefit medical experts in eye disease
diagnostics. [3]
Eman AbdelMaksoud et al. :- In this paper, Researchers developed an ML-CNN
system to diagnose various ocular diseases from color fundus images. They used a
dataset of 45 different diseases and applied augmentation processes to enlarge it. The
system achieved promising performance based on five performance measures and was
compared to other previously proposed systems. In the future, the researchers aim to
establish a large and balanced ML dataset manually to eliminate overfitting. [4]
Harshith N et al. :- In this paper, Researchers used deep neural network model
OpenCV, keras, TensorFlow, pandas, NumPy was developed for eye disease detection.
The model successfully classified between normal and diseased eyes for specific
diseases. The model is cost-effective with a simple interface and provides nearby eye
doctor details for further check-up. The study achieved its objectives of building a deep
learning model for eye disease detection. [5]
Literature Summary
Salam et al. used various CNN models and found VGG-16 to be the most accurate in
classifying ocular diseases from fundus photographs[1]. Kanwal et al. used machine
learning to classify ocular diseases based on symptoms recorded by medical experts, with
tree-based methods performing better than artificial neural networks[2]. Khan et al. used
VGG-19 to classify ocular diseases with high accuracy rates, creating a potential
consumer-friendly categorization system[3]. AbdelMaksoud et al. developed an ML-CNN
system to diagnose ocular diseases from color fundus images, achieving promising
performance measures compared to other proposed systems[4]. Harshith et al. developed a
deep neural network model for eye disease detection, successfully classifying between
normal and diseased eyes and providing nearby eye doctor details[5].
Problem finding
The computational power was required to get the desired results.
Previous researchers have used the same dataset as us, but they only focused on
classifying 2-3 diseases.
Although some of them attempted to classify all diseases, their accuracy was
comparatively lower.
Aims and Objective
The aim of the eye disease classification project is to develop an accurate and efficient
system to diagnose ocular diseases from fundus images.
The objective is to improve early detection and treatment of eye diseases to prevent
vision loss.
Model Design
Data analysis
The data is collected from kaggle. (link :-
https://www.kaggle.com/datasets/andrewmvd/ocular-disease-recognition-odir5k)
There are 6 instances in ocular disease recognition dataset.
This dataset is meant to represent ‘‘real-life’’ set of patient information collected by
Shanggong Medical Technology Co., Ltd. from different hospitals/medical centers in
China.
For balancing the data, we augmented the data using image data generator by
randomizing brightness, shear range, rotation and vertical flipping the images.
For preprocessing the data, we resized the image first then the equalize function is
applied in which we enhanced the contrast of the image and rgb color of the image and
then we applied heatmap which will enhanced the disease images.
Data analysis
Target variable before balancing the dataset:-
Data analysis
Efficient
net Learning 84.25 83.33 82.89
rate:-
0.001
Efficient
net Learning 85.00 82.24 84.49
rate:-
0.0001
Comparative analysis for efficient
netv2:-
Algorithm Parameter Accuracy Precision Recall
s
Efficient Learning
netv2 rate :- 0.01 84.09 82.09 81.30
Efficient
netv2 Learning 82.25 83.33 82.89
rate:-
0.001
Efficient
netv2 Learning 83.22 82.20 81.53
rate:-
0.0001
Comparative analysis for mobile net:-
Algorithm Parameter Accuracy Precision Recall
s
Mobile Learning
net rate :- 0.01 78.90 75.34 76.54
Mobile
net Learning 80.03 83.12 82.50
rate:-
0.001
Mobile
net Learning 81.12 82.35 81.53
rate:-
0.0001
Result for the final model
We get more accuracy in Efficient net but it’s not working that much good on
real world, so we are going with Mobile net model which is not much great on
train data as Efficient net but it works great on real world data.
Accuracy: 86.16
Precision: 84.45
Recall: 85.56
Conclusion & Future Scope