Multi-layer segmentation of retina OCT images via advanced U-net architecture

Optical Coherence Tomography (OCT) is a non-invasive method which can obtain high-definition images of cross section (B-scan) of the retina. By investigating the thickness of different layers of the retina in OCT images, one can diagnose ocular diseases in an early stage. Different algorithms have been proposed for retinal layer segmentation including machine learning techniques and various advanced CNN architectures, which have been developed recently. In this research, segmentation of OCT images is carried out for 9 boundaries, equivalent to segmenting eight retinal layers. We investigate different U-net like structures which can be combined with VGG and ResNet architectures to train models using labelled examples, and accuracy for the predicted retinal layers would be compared. In reducing the complexity of networks, a method is proposed based on the concept of domain decomposition when training a large volume of data on a cloud platform.