Automated segmentation of the in-vivo lamina cribrosa (LC) imaged using 3D optical coherence tomography (OCT)

Abstract

 

Purpose

 

The LC plays an important role in the pathogenesis of glaucoma, yet little work had been done to characterize its microstructures in-vivo. 3D analysis of the LC using manual segmentation is prohibitive due to the extended time required. We aimed to investigate an automated segmentation method for the in-vivo LC imaged using swept-source OCT (SS-OCT).

 

Methods

 

15 subjects (4 healthy, 5 glaucoma suspects, 6 glaucoma) were scanned using a prototype SS-OCT in a 3.5 x 3.5 x 3.64 mm area (400 x 400 x 896 pixels) of the optic nerve head. The region outside the LC was masked and the stack of C-mode slices underwent a median filter, local contrast enhancement and thresholding (Figure). A randomly selected C-mode scan for each subject was independently manually segmented using FIJI by 2 users and compared with the automated method for: (1) segmentation time, (2) sensitivity and specificity assuming pores marked by manual segmentations as the gold standard, (3) pore count, (4) pore area, (5) pore aspect ratio and (6) pore circularity. The measurements (3-6) were compared by computing bias and imprecision using a measurement error model. The automated segmentation and analysis was completed on 8 healthy eyes (Table).

 

Results

 

Average segmentation time was significantly faster using automated segmentation. Average sensitivity and specificity of the automated segmentation for detecting pores was 90.9±9.4% and 87.4±3.3%, respectively. There was no statistically significant difference in bias and imprecision between manual and automated segmentation for all investigated parameters though wide confidence interval were noted for all.

 

Conclusions

 

We herein propose a segmentation algorithm that provides an efficient segmentation of the LC with high sensitivity and specificity and with no statistically significant difference in bias and imprecision from manual segmentation. Wide confidence intervals indicate the inherent difficulty in segmenting the LC. 3D automated segmentation results of healthy LC are consistent with literature values for available parameters.

 

 

Table: LC parameters for healthy eyes.

 

 

Figure: (A) Stacks of cross-sectional images are resampled into (B) C-mode slices (dotted lines on A). (C) Pore (green) and beam (red) segmentations. (D) Overlay of manual (red) and automated (green) segmentation. 3D rendering of the segmented LC beams from the (E) anterior and (F) posterior directions.

 

Keywords: 550 imaging/image analysis: clinical • 627 optic disc