Automated Foveola Localization in Three Dimensional (3D) Optical Coherence Tomography (OCT) Images with Various Macular Pathologies

Abstract

Purpose: : To develop an automated method to determine the foveola location (x, y) in macular 3D spectral domain (SD-) OCT images in either healthy or pathological conditions, for facilitating further analysis around this important landmark.

Methods: : SD-OCT macular scans (Macular Cube 200x200 protocol, 6x6x2 mm; Cirrus HD-OCT; Carl Zeiss Meditec, Inc., Dublin, CA) were obtained from healthy subjects and subjects with various macular diseases (a total of 170 scans from 170 eyes/ 126 subjects; dataset for training [89 eyes from 67 subjects: 30 healthy eyes, 28 with epiretinal membrane (ERM), 37 with macular edema (ME), 16 with age-related macular degeneration (AMD), 17 with macular hole (MH)], and dataset for testing [81 eyes from 59 subjects: 35 healthy eyes, 19 ERM, 31 ME, 13 AMD, 15 MH]). Two ophthalmologists labeled the x and y location of the foveola for each scan independently by going through all x-z and y-z frames. One expert’s labeling was used as the ground truth while the difference between the two experts was used to assess the inter-expert variability. In feature construction, for each sampled location, multi-scale spatially-distributed texture features were computed within the centered context cube/window in the 3D-OCT as well as the OCT en-face image, respectively. Structural Support Vector Machine was trained to discriminatively weigh the features so that the score at each sampled position is inversely proportional to the localization distance. A coarse-to-fine sliding window approach was used to identify the position with the best score. The Euclidean distance between the computed location and the ground truth was used to assess the performance.

Results: : The automated method achieved a mean localization distance compared to the ground truth of 91±53 um (3.03±1.77 pixels) [86±44 um (2.87±1.45 pixels) for healthy and 94±59 um (3.14±1.96 pixels) for diseased eyes]. The inter-expert labeling difference was 54±42 um (1.82±1.39 pixels) [53±41 um (1.78±1.37 pixels) for healthy and 55±43 um (1.84±1.42 pixels) for diseased eyes]. The difference in mean results between the automated method and the inter-expert variability is only 38 um (1.25 pixels), though the difference is statistically significant.

Conclusions: : The proposed automated data-driven method localized the foveola position within 91 um (3.03 pixels) on average from the ground truth, which is well within the anatomical size of the foveola diameter (350um). This method may effectively identify the location of the foveola in healthy or pathological conditions in 3D-OCT scans to facilitate further analysis.

Keywords: image processing • imaging/image analysis: non-clinical • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)