Purchase this article with an account.
Aparna Rao, Rashmi Panda, Niladri Puhan, Debananda Padhy; Automated detection and quantification of retinal nerve fibre layer defects on fundus photography. Invest. Ophthalmol. Vis. Sci. 2017;58(8):687.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To evlaute performance of a computer-aided RNFLD detection alogirthm for efficient automated glaucomatous RNFL changes on fundus images.
A total of 68 fundus images with discernible RNFLD (cases) or glaucomatous changes adjudged by clincian with or without visual field defects were comapred with eyes without RNFLD (controls) and clear media acquired by the which are divided into training set (13 images) and testing set (55 images) Visupac version 4.4.4 camera (FF 450 plus IR Carl Zeiss Ltd. USA) with 5 x magnification and 200, 300, 500 field of views (FOV) were selected. After segmenting and inpainting of blood vessels, Contrast limited adaptive histogram equalization (CLAHE) was applied and candidate RNFLD boundary pixels found by wavelet based local minima analysis of the 1-D intensity profiles extracted from the concentric circles around the optic disc. Novel patch features such as Cumulative Zero Count Local Binary Pattern (CZC −LBP) and Directional Differential Energy (DDE) were applied for boundary pixel classification, Figure 1. For measuring RNFLD angular width, the detected boundary pixels are line fitted using Random Sample Consensus (RANSAC) algorithm. Diagnostic performance of the new algorithm was compared with clinician’s judegement (gold standanrd), two blinded trained graders and Spectral domain OCT. To quantify the closeness of detected RNFLD boundary from the gold standard, mean Euclidean distance (EDmean) was computed.
Agreement between graders for number of RNFD was excellent (ICC 0.86). We tested the proposed new algorithm on 78 RNFLD region boundaries present in 29 of 55 cases. We could detect 61 with 15 FPs using the proposed method which gave an area under the curve of 0.87 using the free response curve with a sensitivity and false positive per image of 78.2% and 0.27, respectively. The maximum value of mean and standard deviation of the detected RNFLD width from the gold standard and existing algorithm are around 10 pixels which implies that the fitted lines are close to the gold standard, Figure 2. The AUC for the algorithm, blinded graders and SD-OCT were 0.92, 0.89&0.78 and 0.83 respectivley. The algirthm picked up 11 small RNFLD missed by OCT and 16 defects missed by the graders.
The proposed method achieves good accuracy as compared to SD-OCT and gold standard which can be used for automated assessment of RNFL changes.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
This PDF is available to Subscribers Only