Purchase this article with an account.
Akram Belghith, Christopher Bowd, Felipe A Medeiros, Robert N Weinreb, Linda M Zangwill; Measurement of BMO-based optic disc area and rim area using Artificial Neural Network Principal component analysis (ANN-PCA) approach applied to 3D spectral domain optical coherence tomography optic nerve head images. Invest. Ophthalmol. Vis. Sci. 2014;55(13):4766.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To propose a new approach for locating the Bruch's membrane opening (BMO) and estimating the BMO-based optic disc area and rim area of 3D Spectralis SD-OCT images (Heidelberg Engineering, 48 enhanced depth imaging EDI radial B-scans centered on the optic nerve head).
We formulated the detection of the BMO as a missing data problem where we jointly estimate the noise hyper-parameters and the segmented image. To deal with the overlapping of the Bruch's membrane and the retinal pigment epithelium layers due to the poor image resolution, we propose the use of an image deconvolution approach that consists of assigning to each layer a specific shape (or filter) and then estimating its hyper-parameters. To estimate the BMO-based optic disc size and rim area, we propose the use of a new regression method based on the Artificial Neural Network Principal component (ANN-PCA) analysis, which allows us to model irregularity in the BMO estimation due to scan shifts and/or poor image quality. Axial length corrected rim area measurements using this approach were compared to those from Cirrus SDOCT (Optic disc Cube 200x200, Cirrus OCT). The diagnostic accuracy of rim area, and rim to disc area ratio was also compared to Retinal Nerve Fiber Layer (RNFL) thickness measurements for glaucoma detection. Glaucoma diagnostic accuracy (area under receiver operating characteristic (AUROC)) was estimated using 105 glaucoma and 100 healthy eyes.
The BMO-based disc area and rim area estimations with the new ANN-PCA approach using Spectralis EDI radial scans were similar to Cirrus measurements (Figure 1). The correlations between the two devices for the BMO-based disc area was R2 = 0.845 (p <0.001), for the rim area was R2 = 0.783 (p <0.001) and for the RNFL thickness was R2 = 0.848 (p <0.001). The diagnostic accuracy of RNFL tended to be better than rim area and rim to disc ratio but varied by instrument (Table 1).
Our proposed automated approach for estimating the BMO-based disc area and rim area measurements of 3D Spectralis SD-OCT images provides similar results to those from Cirrus SD-OCT. AUROC results suggest that RNFL may still be the measurement of choice for discriminating between glaucoma and healthy eyes.
This PDF is available to Subscribers Only