This repository provides source code, submitted papers and demo for Diabetic Retinopathy: Segmentation, Grading and Localization with IDRiD dataset.Our method won the 1st place in Fovea Localization with overall 3rd place in the Localization sub-challenge of IDRiD Grand Challenge. Besides, we secured 5th rank in Segmentation of Hard Exudates.

We developed a data-driven method for automatic detection of retinal lesions based on their characteristics in fundus images to treat Diabetic Retinopathy. First we preprocessed retinal images to reduce image noise and did several data augmentations to make data variety more richer and distinct. After that, we segmented them using a model with UNet architecture and weighted cross entropy loss. The Unet architecture includes a shrinking path to capture context of surrounding and a symmetric expanding path that enables accurate localization. We improved and extended this architecture such that it works with very few training images and produces more accurate lesion masks. We have trained four distinct binary models for the four lesions - hard exudates, hemorrhages, microaneurysms and soft exudates.

The retina images for disease classification were obtained from Kaggle dataset. The labels were provided by clinicians who rated the severity of diabetic retinopathy in each image on a scale of 0-4. The images were preprocessed, downsampled and augmented before feeding them to a 50 layer deep ResNet. Our networks achieved very competitive Kappa score of 0.74 on Kaggle Private Leaderboard, along with sensitivity of 82% and specificity of 84%.

The model consisted of two subsequent approaches to do localization. The initial model was a Convolution based model to get a tentative position of the Fovia and the Optical Center. The next model worked on a patch around the predicted value of the previous model to improve the accuracy. The motivation for sucha step was to first get a global view of the image and then focus on a local area for predicting the coordinates. The first model is a standard CNN while the second model implements UNet architecture.

The repository also provides short papers submitted to IEEE ISBI as part of the IDRiD Grand Challenge.

We also developed a Django webserver with a basic user interface to upload and test retinal images on our models. The Django site was hosted on our workstation. View demo video: https://apopli.github.io/dr_demo.mp4