Abstract
Purpose :
The correct detection and localization of the Irido-Corneal Angle (ICA) in gonioscopic images are of paramount importance for automatic alignment of the NIDEK GS-1 device (NIDEK CO., LTD Japan) on the patient's eye. The intrinsic variability of the anatomical structures, the effect of pathologies, and the possible presence of eye implants or imaging artifacts (air-bubbles in the gel, focusing issues, etc.) make this task very hard for classical image processing techniques. The aim of this work is to investigate an alternative Deep Learning (DL) based algorithm for addressing this problem.
Methods :
Over 10k 1280x960 RGB images have been collected during clinical evaluation of the GS-1 device. In most of them the ICA is visible; in the others, acquired in misaligned conditions, no eye structures are actually recognizable. A subset of 3412 images has been manually annotated indicating the presence of any eye structure, the iris shade (Dark/Light/NotSure), the ICA visibility (i.e. visibility of both iris and sclera), and the angle classification (Closed/Open/NotSure). Where applicable, a Reference Point (RP) has been manually placed over the intersection of the trabecular meshwork (if visible) or of the iris-sclera limit with the image median. This dataset has been split into two sets of 1019 and 2393 images with the same statistics with respect to the above classification criteria. Cross-correlation has been avoided by constraining all images referring to the same patient to belong to the same subset. A Convolutional Neural Network (CNN) has been designed for detecting whether the ICA is present or not and, if present, for locating the RP in the input image. The first subset has been used for training the CNN, the second for validating the results.
Results :
The proposed approach outperforms the classical one: RP mislocation and false positive rates decrease from 6.8% to 2.5% and from 10% to 3.7%, respectively; the overall success rate (correct detection and localization) increases from 79% to 91%.
Conclusions :
The features computed by the trained CNN turn out to be very reliable for ICA detection and RP localization. Hence, the GS-1 auto-alignment procedure should greatly benefit from the introduction of this algorithm for offering a better exam acquisition experience.
This is a 2020 ARVO Annual Meeting abstract.