Diabetic Retinopathy (DR) is a diabetes complication that affects the eyes. It is caused by damage to the blood vessels of the light-sensitive tissue at the back of the eye (retina). DR can lead to blindness if not detected and treated early. Early diagnosis and timely treatment are crucial to prevent vision loss.

1.Business Problem

2.Visualization of Images

3.Image Processing

4.TSNE Visualization

5.Data Augmentation

6.Train Validation Split

7.Resize and save

8.Image Pre Processing and save

9.ResNet50 Models

This case study is based on a Kaggle Competition conducted 7 months back based on a dataset by Aravind Eye Hospital in India to detect a form of Blindness called Diabetic Retinopathy. https://www.kaggle.com/c/aptos2019-blindness-detection/overview is the link to the Kaggle competition.

The goal here is to Build an Image Classification Model which can take a look at the images and classify the image into one of the 5 classes (0,1,2,3,4). This Image Classification Model can accelerate the process of Blindness Detection in Patients. Currently Doctors review the Image and classify it into one of the classes -

0 - No DR

1 - Mild

2 - Moderate

3 - Severe

4 - Proliferative DR

Training data contains 3662 images, test data contains 1928 images

• Gaussian Blur

• Gaussian Blur with Circular Cropping

ACCURACY

• Train Cohen Kappa score: 0.980

• Train Accuracy score : 0.965

Evaluate Model on Test Data

• Test Cohen Kappa score: 0.904

• Test Accuracy score : 0.839

+-------+----------------------------+----------------------------+----------------------+-------------------------------+-----------+----------+ | S.No. | ResNet50 Model | Image Processing | Data Augmentation | Hyperparameters(BS,Opt,lr,ep) | Train QWK | Test QWK | +-------+----------------------------+----------------------------+----------------------+-------------------------------+-----------+----------+ | 1 | R-P-D-p(0.5)-D-p(0.5)-S(5) | -- | Hor Flip,Scale 1/255 | (4,'Adam','1e-4',7) | 0.912 | 0.905 | | 2 | R-P-D-p(0.5)-D-p(0.5)-S(5) | Circle Crop, Gaussian Blur | Hor Flip,Scale 1/255 | (4,'Adam','1e-4',7) | 0.98 | 0.904 | +-------+----------------------------+----------------------------+----------------------+-------------------------------+-----------+----------+

Python 3.8

Flask

Tensorflow

Keras

OpenCV

Flask

JavaScript

CSS

HTML