Skip to main content Link Search Menu Expand Document (external link)

Deep Retinal Image Understanding

K.K. Maninis, J. Pont-Tuset, P. Arbeláez, and L. Van Gool

Approach of the paper

Specialized task of a base segmentation network to specific tasks like - blood vessel segmentation and optical disk segmentation - in the retina

  • basenet - VGGnet
  • task specific specialized convolutional layers - inspired by GoogLeNet

Specifics of the network

  • specialized layers add supervision to multiple internal layers.
  • it produces feature maps in k different channels which are resized to original image size and concatenated thereby creating fine to coarse feature maps
  • a last convolutional layer linearly combines the feauture maps created by specialized layer to regressed result.
  • class balanced cross entropy to train the network to detect contours.
  • first 4 finer feature maps for blood vessel segmentation and last 4 coarse feature maps for optic disk segmentation

Sample reporduced results