loading page

A deep learning based approach to segment exudates in retinal fundus images using Recurrent Residual U-Net
  • +1
  • Ninad Mehendale ,
  • Kush Vora ,
  • Divesh Thakker ,
  • Darshil Mehta
Ninad Mehendale
K. J. Somaiya College of Engineering, K. J. Somaiya College of Engineering, K. J. Somaiya College of Engineering

Corresponding Author:[email protected]

Author Profile
Kush Vora
Author Profile
Divesh Thakker
Author Profile
Darshil Mehta
Author Profile


Diabetic Retinopathy (DR) is a severe medical concern, and early detection of DR can prevent blindness. Manually examining fundus images to detect lesions is time-consuming and requires expertise. Segmenting exudates from fundus images using Deep Learning methods provides an automated high precision solution to this problem. Our work proposes the Residual Recurrent U-Net (R2 U-Net) for segmenting exudates in retinal fundus images. The network consists of recurrent units with feedback connections leveraging local spatial information from neighboring pixels. There are multiple stacked recurrent units in each layer of the U-Net. Residual skip connections were introduced between different layers of the network to train deeper architectures. The model was trained on the publicly available IDRiD dataset and tested on IDRiD, E-Ophtha, and DiaretDB1 dataset. The metrics were computed at the pixel, exudate (or lesion), and image level. The model achieved a state-of-the-art 93.20% sensitivity and 99.80% specificity on the E-Ophtha at the exudate level, which is better than existing literature. An automated DR detection system would help not only people but also Ophthalmologists to treat Diabetic Retinopathy effectively.