Retinal layer delineation through learning of tissue photon interaction in optical coherence tomography

Abstract

Optical coherence tomography (OCT) is widely used in Ophthalmology for visualizing retinal layers and cornea width profiles which are indexed to various pathologies. Image processing algorithms have been deployed to automatically delineate various layers for autonomous width profiles computation. Classical OCT segmentation algorithms are anchored around gradient and its derivatives. Such pixel information can vary from subject to subject so prior approximations of different layers are imposed in the name of regularization. Regularization restrains the generalization capability of the model which is crucial for segmentation in pathological subjects. The presented approach aims to model signal interactions with a tissue type through back scattered signal characteristics. During prediction phase, given an unknown interaction, the model estimates the probability of signal being backscattered from a tissue type. Sensor's ballistic models are employed to estimate the uncompressed signal statistics. These models have been experimented and evaluated on 5000 AMD (age-related macular degeneration) and 5000 normal B-scans from Duke OCT dataset. With two percent training data the delineation results are comparable to graph based approaches. The anterior retina, retinal pigment epithelium (RPE) and posterior retina are identified with sensitivity of 0.87, 0.84 and 0.91 respectively.