Segmentation of Diabetic Retinopathy Images

This directory comprises code to preprocess, train, test and evaluate the model. The implementation is based on Tensorflow.

Requirements

You need to have the following python packages installed:

• tensorflow
• opencv
• numpy
• scipy
• sklearn
• matplotlib
• PIL
• pandas

You need to have downloaded the IDRiD training dataset.

How to use

Set paths

We need to know where you saved the data and where you wish to save the results. Go into the respective python scripts and modify the paths to suit your setup.

Data preprocessing

The ground truths are thresholded to levels 0 and 255.

python gt_binary.py

Since the retinal image is of a very high resolution, smaller-sized overlapping patches are extracted from the training images and masks. This is done while maintaining the ratio of patches with lesions do those without lesions.

python extract_patches.py

The patches are separated into training and validation data, and the respective file names are written in train.csv and test.csv.

python traintest_split.py

Network training

Now train the UNet model on the dataset of patches. In order to also validate on test split while training

python train.py --epochs=100

To train without validation

python train_100.py --epochs=100

Network testing

To test the trained model on a single patch

python test.py

To test on an entire image

python test_image.py

Generate probability maps for the dataset of retinal images using the trained model

python prob_images.py

Generate segmented output masks from the probability maps

python test_images.py

Evaluation

Calculate sensitivity and precison values for individual images, plot FROC curve, compute average statistics and precision score

python sn_ppv.py
python froc.py
python stats_avg.py
python score.py