Abstract
Purpose :
The choroid, known to be affected in conditions such as age-related macular degeneration (AMD), has become more accessible with modern imaging techniques, although quantifying its structure can be time consuming and prone to high inter-observer variability. Automated feature analysis of the choroid may overcome these issues; we therefore performed a prospective cross-sectional study to investigate the potential of automated classification of choroidal images by AMD disease stage. The discriminatory image features are identified, and their properties and potential structural origins are evaluated.
Methods :
Long-wavelength (1040 nm) optical coherence tomography (OCT) volume scans were acquired from eyes with early AMD (n=25), neovascular AMD (n=25), and healthy age-similar controls (n=25). Gabor filters of varying spatial frequency and orientation were convolved across the image to extract textural information from the choroidal region of each b-scan. These image features were used to train and test a random forest classifier to coarsely sort images based on disease severity. The relative importance of each filter in the classifier was established by reclassifying with each filter in turn removed from analysis, and calculating the difference in the accuracy obtained (permuted delta error, normalised from 0 to 1; δ).
Results :
A classification accuracy of 89.2% and 60.0% was achieved with 10-fold and leave-one-eye-out cross-validation respectively. Despite the relatively small sample size, this exceeds chance for a 3-class problem. The lowest spatial frequency filters were of highest importance to the classifier (δ=0.81), whilst the mid-frequency filters were of lowest importance (δ=0.48). The vertical gratings were also of highest importance (δ=0.88), particularly for the higher spatial frequency filters.
Conclusions :
Feature analysis of the choroid can be used to differentiate between AMD disease stages. This is based on the assumption that choroidal changes in AMD follow the retinal changes observed. Care should be taken with interpretation of the vertical components, as these may be sensitive to shadow artefacts from overlying features such as drusen, and therefore not choroidal in origin. With appropriate respect given to the potential limitations of this technique, feature analysis provides a fast, automated quantification of choroidal structure from OCT images.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.