Logo MachVis


RASNet: Semantic Segmentation of Retinal Anatomical Structures Using Deep Neural Network
M Badar, M Shahzad, Q Khan, M M Fraz

Diabetic retinopathy, glaucoma and macular degeneration are the leading causes
of blindness in the world. An automated periodic screening system can reduce
the threat to visual acuity. In this regard, accurate segmentation of retinal
anatomical structures and pathologies is of primary importance and the fun-
damental step in the development of automated retinal diagnostic/screening
systems. In this context, we propose a novel end-to-end framework for simulta-
neous segmentation of multiple retinal pathologies (i.e., exudates, hemorrhages,
and cotton-wool spots) and landmarks (i.e. retinal blood vessels and optic
disc) using an encoder-decoder based fully convolutional deep neural network
architecture named as RASNet (Retinal Anatomical Structures segmentation
NETwork). The extraction of retinal landmarks and pathologies has been mod-
elled as a semantic segmentation problem which enabled us to obtain pixel-level
labels for each class. For the purpose of evaluation, a new public retinal im-
age dataset (named as RAS-DB) is created, where the pixel-level multi-class
annotations are done by the ophthalmologist.The proposed RASNet is evalu-
ated on DRIVE, STARE, CHASE DB1 and RAS-DB datasets, and it achieved
state of the art performance in retinal anatomical structure segmentation using
standard evaluation metrics.


The data-set and code will be made public after article publication.