eduzhai > Applied Sciences > Engineering >

Improving Lesion Segmentation for Diabetic Retinopathy using Adversarial Learning

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 12 pages

Abstract: Diabetic Retinopathy (DR) is a leading cause of blindness in working ageadults. DR lesions can be challenging to identify in fundus images, andautomatic DR detection systems can offer strong clinical value. Of the publiclyavailable labeled datasets for DR, the Indian Diabetic Retinopathy ImageDataset (IDRiD) presents retinal fundus images with pixel-level annotations offour distinct lesions: microaneurysms, hemorrhages, soft exudates and hardexudates. We utilize the HEDNet edge detector to solve a semantic segmentationtask on this dataset, and then propose an end-to-end system for pixel-levelsegmentation of DR lesions by incorporating HEDNet into a ConditionalGenerative Adversarial Network (cGAN). We design a loss function that addsadversarial loss to segmentation loss. Our experiments show that the additionof the adversarial loss improves the lesion segmentation performance over thebaseline.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...