Feature Definition and Selection for Epiretinal Membrane Characterization in Optical Coherence Tomography Images

Abstract

[Absctract]: Optical Coherence Tomography (OCT) is a common imaging technique for the detection and analysis of optical diseases, since it is a non invasive method that generates in vivo a cross-sectional visualization of the retinal tissues. These characteristics contributed to the use of OCT imaging in the analysis of pathologies as, for instance, vitreomacular traction, age-related macular degeneration or hypertension. Among its applications, OCT imaging can be used in the detection of any present epiretinal membrane section in the retina, a critical issue to prevent further complications caused by this pathology.

This work analyzed the main characteristics of the epiretinal membrane to define a complete and heterogeneous set of intensity and texture-based features. Those features were studied using representative selectors, as Correlation Feature Selection (CFS) and Relief-F, to identify the optimal subsets that offer the higher discriminative power. K-Nearest Neighbor (kNN), Naive Bayes and Random Forest were finally tested in a method for the automatic detection of the epiretinal membrane in OCT images. Previous works do not focus on automatic procedures and, on the contrary, depend on manual markers or supervised detections, while our method improves significantly this task by automating the search of the region of interest and the classification of the pixels belonging to that area.

The methodology was tested using a dataset of 129 OCT images. 120 samples were equally obtained from those scans, featuring both zones with and without epiretinal membrane. The best results were provided by the Random Forest classifier that, using a window size of 15 pixels, a quantity of 13 histogram bins and 28 features, achieved an accuracy of 93.89%.