Three Dimensional Fully Convolutional Networks for Segmentation of Optical Coherence Tomography Images in Neurodegenerative Disease

F. Kiaee, H. Fahimi, R. Kafieh, A. U. Brandt, H. Rabbani

Abstract:

Optical coherence tomography (OCT) is an important method for visualization
and quantification of intra-retinal layers. OCT-derived measures of retinal layers
support investigating the role of afferent visual pathway degeneration in neurode-
generative diseases like multiple sclerosis (MS). Therefore, accurate, robust and
repeatable segmentation of retinal layers is of interest in such applications. In this
paper, a novel 3D fully convolutional deep architecture is proposed for automated
segmentation of retinal layers. For this purpose, 3D convolutions explore spatial
and inter-frame dimensions to extract features. The proposed network uses a set of
convolution and subsampling layers in an alternating fashion to learn a hierarchy
of shrinking 3D feature maps (encoder stage). The encoder is then followed by
multiple convolution and upsampling blocks enlarging the feature map to the size
of original input image for semantic segmentation (decoder stage). The proposed
framework was validated on 3D OCT scans of healthy subjects captured by a
Topcon 3D OCT device. We contrast the ensemble results with the Deep-Net-2D
and Graph-DP methods and observe a significant increase of 6% in the Dice metric
for two layers and consistent improvements across the retinal layers. Despite the
strategies used for dealing with the class imbalance, contour error values are rather
inferior for two layers, but still promising for most of the classes. The results are
promising for further application of the approach in neurodegenerative diseases.

Author Affiliation: Isfahan University of Medical Sciences, Isfahan, IRAN, Charité - Universitätsmedizin Berlin, Berlin, German

Keywords: Optical coherence tomography, fully conventional deep architecture, neurodegenerative diseases