Non-invasive measurement of choroidal volume change and ocular rigidity through automated segmentation of high-speed OCT imaging

Abstract

We have developed a novel optical approach to determine pulsatile ocular volume changes using automated segmentation of the choroid, which, together with Dynamic Contour Tonometry (DCT) measurements of intraocular pressure (IOP), allows estimation of the ocular rigidity (OR) coefficient. Spectral Domain Optical Coherence Tomography (OCT) videos were acquired with Enhanced Depth Imaging (EDI) at 7Hz during ~50 seconds at the fundus. A novel segmentation algorithm based on graph search with an edge-probability weighting scheme was developed to measure choroidal thickness (CT) at each frame. Global ocular volume fluctuations were derived from frame-to-frame CT variations using an approximate eye model. Immediately after imaging, IOP and ocular pulse amplitude (OPA) were measured using DCT. OR was calculated from these peak pressure and volume changes. Our automated segmentation algorithm provides the first non-invasive method for determining ocular volume change due to pulsatile choroidal filling, and the estimation of the OR constant. Future applications of this method offer an important avenue to understanding the biomechanical basis of ocular pathophysiology.

Keywords: (170.3880) Medical and biological imaging; (170.4460) Ophthalmic optics and devices; (170.6935) Tissue characterization.

Fig. 1

Automated segmentation of retinal layers of interest from OCT images. The image in A is a typical frame from the video series. A) A-scans where the choroid is absent (highlighted in green) are discarded from all frames. B) Segmentation of the outmost layers of the retina: RVI is indicated in green, anterior RPE in red, and posterior RPE in blue. The A-scans are shifted so that the blue layer appears flattened (C).

Fig. 2

A) The uncompensated sub-RPE region. B) The same region, compensated and contrast enhanced, to which the oriented gradient algorithm will be applied. Overlaid is an example disk of radius 30 pixels. The red and green regions correspond to their respectively coloured histograms (D). C) The oriented gradient image, composed of the combination of the X² images of different scales and orientations, as described. The heat map shows pixels which are very likely to lie on a boundary. Even weak boundaries can be detected while excluding noisy regions with this method. E) Overlay of the oriented gradient image (heatmap), node locations (yellow x’s), and the CSI found using these two inputs to the graph search (redline), onto the flattened B-scan. The green dashed line shows the limit of 585 µm below the Bruchs beyond which nodes are discarded. F) The original B-scan overlaid with the RPE (blue), CSI (yellow) and the mean RPE-CSI distance or CT (red dotted line). This distance CT is what is tracked from frame to frame.

Fig. 3

Spectrum analysis of CT fluctuations in time. A) Top: Oximeter signal. Bottom: Raw fluctuations of CT versus time (black). Overlaid in red is the band-pass filtered CT signal (red). B) Frequency spectrum of the oximeter signal (top), and CT signal (bottom), where the offset component has been omitted. The filtered frequency band for the CT spectrum is shown in red..The dashed blue lines indicate the two first harmonics of the measured heart rate which are observed in both signals.

Fig. 4

Segmentation results for the proposed method (blue), Tian’s method (red) and manual tracings (green). A set of 25 OCT images was manually segmented by 5 independent specialists and used for benchmarking our method. A, B) Examples illustrating the performance. C) We calculated the average manually segmented CSI and compared the performance of each specialist and both automated methods. Histograms of deviation from the mean trace for every A-scan of all images are shown as violin plots

Fig. 5

Ocular rigidity measurements. A) Correlations of OR with OPA and AL. B) Reproducibility results (4 subjects). Along with ICC value, its lower and upper bonds for alpha = 0.05, are indicated.