Analysis of the choroid is important to assess diseases such as glaucoma,age-related macular degeneration(AMD),serous chorioretinopathy and choroidal melanoma,which accompany choroidal changes [5].A common imaging modality to visualize the retinal layers and the choroid is optical coherence tomography(OCT).An OCT scan comprises of hundreds of slices(called B-scans)which makes manual analysis of a scan highly labour intensive.This is especially of a concern when most of the examined scans turn out to be normal.Moreover,manual examination can be highly subjective.In order to save the labour and avoid the subjectivity involved in manual analysis,automatic analysis of the choroid can be highly beneficial for clinical applications. The first step in automatic analysis of the choroid is to segment it in OCT B-scans(referred to henceforth as'images').Even though there has been extensive research in automatic retinal layer segmentation,only a few works segment the choroid.Existing choroid segmentation methods include graph search [13],statistical models [9] and other image processing methods [2,5].Another possible method is to use deep learning which has shown better performance than conventional machine learning techniques for many image segmentation tasks.Deep networks can learn task-specific high-level discriminative features.Sui et al.[12] proposed to use a convolutional neural network(a deep learning tool)to provide a probability map which was used by graph search to segment the choroid.A similar approach was used by Alonso-Canero et al.[3] who also performed an initial segmentation of the choroid using CNN and refined it using graph search.Hassan et al.[8] segmented the retinal layers and the choroid using a tensor-based method and refined it using CNN.
展开▼
