Automatic Detection of Diabetic Retinopathy with Deep Learning

Detection of fundus in eyeball images for early detection of Diabetic Retinopathy using Deep Learning and classifying them into Not Present, Mild, Moderate, or Acute.

Running the script

Make sure you have the following libraries installed:

• tqdm
• tensorflow
• keras
• matplotlib
• numpy
• pandas
• requests
• colorama
• scikit-learn
• opencv

by running the following command:

pip install -r requirements.txt

Next, run the main script using the command:

python main.py

The main script downloads DIARETDB0 from the website if not already available in your current directory, and unzips it. A series of functions are called to normalise the images, annotate them for classes using a scoring function, create a dataframe, and finally pass it to our Convolutional Neural Network for training.

Training history can be visualised on Tensorboard using the following command:

tensorboard --logdir = ./logs

Running the script for 200 epochs gave us a baseline accuracy of 80%. Currently, we are working on improving our model to increase accuracy scores on the test dataset.

Scoring Function

The annotations are stored as .dot files in the downloaded data directory: diaretdb0_v_1_1/resources/images/diaretdb0_groundtruths/. The files contain annotations in the format: redsmalldots hemorrhages n/a n/a n/a which accounts for 5 levels of fundus in Diabetic Retinopathy images, in ascending order of severity.

We convert this annotation into a 5-bit binary number and then find its decimal equivalent as follows:

• n/a is accepted as False or 0
• Everything else is regarded as True or 1

Our scoring and class function adhere to the following rules:

Lower Limit	Upper Limit	Class
0	0	0
1	1	1
2	12	2
13	32	3

Example:

Image	2^0	2^1	2^2	2^3	2^4	Score	Class
image113.dot	n/a	n/a	n/a	n/a	n/a	0	0
image060.dot	redsmalldots	n/a	n/a	n/a	n/a	1	1
image081.dot	redsmalldots	haemorrhages	n/a	n/a	n/a	3	2
image011.dot	redsmalldots	haemorrhages	hardexudates	softexudates	n/a	31	3

Visualising the Dataset

To visualise the dataset images, run:

python scripts/feature_extraction.py --img <IMAGE>

where <IMAGE> can have values as image001.png, image002.png till image130.png. Running the script should give you the results such as the following:

Example: python scripts/feature_extraction.png --img image011.png

The blood vessels have been separated from the exudates.

Tasks:

• Image Preprocessing
• Train Model
• Annotation GUI using Tkinter
• Unsupervised Clustering of images

Group Members:

• Sumit Saha - sumit.saha.acad@gmail.com
• Ishan Bhattacharya - ishan98b@gmail.com