Characterization of sources and methods for optic disc head location determination in fundus images

Abstract

Purpose : Scanning laser ophthalmoscope (SLO) fundus images obtained during optical coherence tomography (OCT) are manually centered on the optic disc center (ODC) during initial manual centration by qualified study personnel (termed M0). In this investigation, the ODC was determined via centroid calculation based on manual determination of the circumference vertices of the optic nerve head (termed M1), and automatically via imaging algorithms (termed A1). We aim to characterize the accuracy of M0 and A1 determination with respect to M1. Due to manual detection of the optic disc (OD) edge and because random vertex errors partially cancel during centroid calculation, M1 is assumed to be most accurate.

Methods : Deidentified SLO images from the Leipzig Research Centre for Civilization Diseases - LIFE Adult study were employed and M0 centers were extracted from the Spectralis OCT, Heidelberg Engineering machine. M1 was calculated as the centroid of 16 manually determined points at equiangular radial directions along the circumferential edge of the OD for 317 images, as guided by an annotation software coded in R. The detection of A1 was performed in two steps for each image: 1) binarize the original image via blood vessel detection; extract the largest binary object; find a convex blob whose center point designates the center of a region of interest (ROI) for step 2) use the ROI on the original image; erase the blood vessels via grayscale closing operation followed by a Gaussian smoothing filter; apply the circular Hough transform to find the circle radius and the center A1 from the best OD fit.

Results : Fig. 1 shows horizontal and vertical M0-M1 difference distribution fits and Fig. 2 shows horizontal and the vertical A1-M1 difference distribution fits. Using the Akaike Information Criterion, we find that the T family distribution fits the data best. Both figures display the three quartiles (Q1-25%, Q2-50%, and Q3-75%) of the distributions. The accuracy for both the manual and automatic (in the vertical direction) methods is within 0.01 mm. The horizontal accuracy of the automatic method is 0.03 mm. The interquartile range is between 0.08 - 0.11 mm.

Conclusions : We implemented and investigated an automatic imaging method for ODC detection with comparable accuracy to manual centration. This is clinically meaningful as now large sets of images can be automatically analyzed.

This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.

 

Figure 1.

 

Figure 2.