Skip to main navigation Skip to search Skip to main content

Automatic fundus image segmentation for diabetic retinopathy diagnosis by multiple modified U-nets and segnets

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Abstract

Diabetes mellitus leads to damage of the retina by a high blood sugar level. This disease is called diabetic retinopathy (DR), and it is one major cause of blindness among working-aged people. DR affects about 80% of patients who have had diabetes for twenty years or more. The longer a period of diabetes is, the higher the risk of developing DR is. In order to prevent the blindness caused by DR, accurate DR diagnosis from a retinal fundus image is important. Recently, deep learning techniques play significant role in the field of computer vision. When we apply deep learning to segmentation of abnormal parts in fundus images, two major problems arise. One is that the number of available data is insufficient to train a deep neural network. The other is that the sizes of the abnormal parts are quite different depending on the type of the disease, which leads to low segmentation accuracy of small diseases. These two problems make the fundus image segmentation challenging. In this paper, we propose a segmentation method using multiple deep neural networks. To train the deep neural networks from a small number of data, we use data augmentation as preprocessing and adopt the Dice coefficient with the binary cross entropy as a loss function. Moreover, to improve the segmentation accuracy of small diseases, e.g., microaneurysms, we construct one individual network for each type of the disease. In experiments, the networks are trained from IDRiD dataset and tested for MESSIDOR dataset. We compare and discuss the accuracy of the proposed method with modified U-Nets and SegNets.

Original languageEnglish
Title of host publication2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2019
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages1582-1588
Number of pages7
ISBN (Electronic)9781728132488
DOIs
Publication statusPublished - 2019 Nov
Externally publishedYes
Event2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2019 - Lanzhou, China
Duration: 2019 Nov 18 → 2019 Nov 21

Publication series

Name2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2019

Conference

Conference2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2019
Country/TerritoryChina
CityLanzhou
Period19/11/18 → 19/11/21

UN SDGs

This output contributes to the following UN Sustainable Development Goals (SDGs)

  1. SDG 3 - Good Health and Well-being
    SDG 3 Good Health and Well-being

ASJC Scopus subject areas

  • Information Systems

Fingerprint

Dive into the research topics of 'Automatic fundus image segmentation for diabetic retinopathy diagnosis by multiple modified U-nets and segnets'. Together they form a unique fingerprint.

Cite this