Dense Correlation Network for Automated Multi-Label Ocular Disease Detection with Paired Color Fundus Photographs

In ophthalmology, color fundus photography is an economic and effective tool for early-stage ocular disease screening. Since the left and right eyes are highly correlated, we utilize paired color fundus photographs for our task of automated multi-label ocular disease detection. We propose a Dense Correlation Network (DCNet) to exploit the dense spatial correlations between the paired CFPs. Specifically, DCNet is composed of a backbone Convolutional Neural Network (CNN), a Spatial Correlation Module (SCM), and a classifier. The SCM can capture the dense correlations between the features extracted from the paired CFPs in a pixel-wise manner, and fuse the relevant feature representations. Experiments on a public dataset show that our proposed DCNet can achieve better performance compared to the respective baselines regardless of the backbone CNN architectures.
  • IEEE MemberUS $11.00
  • Society MemberUS $0.00
  • IEEE Student MemberUS $11.00
  • Non-IEEE MemberUS $15.00
Purchase

Videos in this product

Dense Correlation Network for Automated Multi-Label Ocular Disease Detection with Paired Color Fundus Photographs

00:11:15
0 views
In ophthalmology, color fundus photography is an economic and effective tool for early-stage ocular disease screening. Since the left and right eyes are highly correlated, we utilize paired color fundus photographs for our task of automated multi-label ocular disease detection. We propose a Dense Correlation Network (DCNet) to exploit the dense spatial correlations between the paired CFPs. Specifically, DCNet is composed of a backbone Convolutional Neural Network (CNN), a Spatial Correlation Module (SCM), and a classifier. The SCM can capture the dense correlations between the features extracted from the paired CFPs in a pixel-wise manner, and fuse the relevant feature representations. Experiments on a public dataset show that our proposed DCNet can achieve better performance compared to the respective baselines regardless of the backbone CNN architectures.