Purchase this article with an account.
Simone Pajaro, Lorenzo Cappellari, Andrea De Giusti, Anna Paviotti; Trabecular meshwork detection in automated gonioscopy. Invest. Ophthalmol. Vis. Sci. 2018;59(9):5867.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
In traditional gonioscopy the anterior chamber angle is subjectively assessed by manual inspection of the Trabecular Meshwork (TM) and nearby eye structures. NIDEK GS-1 (NIDEK CO., Japan) acquires digital true color images of the irido-corneal angle. Automatic detection of the TM in these images is crucial for Automated Gonioscopy (AG) and for quantitative measurements on the angle features. The aim of this study is to evaluate different approaches to automatic TM detection.
A dataset of 318 images was collected during the clinical evaluations of the GS-1 device. The 1280x960 RGB images were acquired during the alignment phase and refer to the Temporal, Superior, Nasal, and Inferior quadrants. Using a specifically designed SW tool, two experts manually identified the TM location on every image as the intersection between the TM line and the image median. Besides, the operators classified the iris shade (Dark/Light/Hazel) and the TM status (Present/NotPresent) on all images. On 189 images the iris shade was classified as Dark, on 68 images as Light, and on 61 as Hazel. On 311 images the TM was deemed to be Present, and on 7 NotPresent. Three different TMD algorithms were run on the dataset. The first (TMD1) combines multi-channel segmentation with intensity profile. The second (TMD2) identifies candidate TM points by thresholding and then fits them with an outlier-robust iterative method. The third (TMD3) uses a statistical approach to detect the candidate TM points.
The TM detection was considered successful either if the Euclidean distance between the estimated and the manually identified TM points was within 100 pixels, or the detection failed on an image where the TM was not present. The success rate was 89%, 94%, and 93% for TMD1, TMD2, and TMD3, respectively. The wrong detection rate (TM detected in a wrong position or where not present) was 11%, 3%, and 4%. The missed detection rate (TM not detected where present) was 0%, 3%, and 3%.
Automatic TM detection is a difficult task, due to the high variability in the morphology and pigmentation of pathological eyes. Despite that, all algorithms successfully detected the TM on about 90% of the images with enough accuracy to perform AG. The performance should be improved in terms of the wrong detection rate. In this respect, TMD2 and TMD3 algorithms show the most promising results.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
Success case for all the three analyzed algorithms.
Wrong/missed detection case.
This PDF is available to Subscribers Only