Abstract
Purpose :
Automatic measurement of the anterior chamber (AC) using OCT plays an important role in both ophthalmic disease diagnosis and treatment, such as glaucoma risk assessment and intraocular lens (IOL) calculation. However, due to the discontinuity of the iris and interference of eyelid and eyelashes, accurate segmentation of anterior chamber is challenging. In this study, we developed a deep learning (DL) method to segment AC for SS-OCT.
Methods :
A prototype algorithm was developed to segment the anterior and posterior corneal and iris surfaces of the anterior chamber scans. We used the prototype segmentation to generate the ground truth for input images, followed by manual grading of the images to exclude segmentation errors (Fig 1). A total of 10,950 images from 64 volumes (14 subjects) with PLEX® Elite 9000 (ZEISS, Dublin, CA) with 6x12x18 mm cube scans (3072 pixels x 400 A-lines x 600 B-scans) were used to generate two separate models for initial cornea and iris segmentation (Bagherinia et al., IOVS June 2023, Vol.64, 1115). Augmentations including geometrical and photometric were applied to increase the training set. A postprocessing algorithm adjusted the initial segmentations from the DL models. The performance of the algorithm was evaluated using 184 images randomly selected across the 6x12x18 mm cube data from 9 scans (5 subjects) and reported by human manual grading. The segmentation was reviewed by human graders who rated the level of segmentation quality as being acceptable or failed. The success rate of each segmented surface was reported.
Results :
Fig 1 shows examples of 6x12x18 mm OCT B-scan images and the corresponding ground truth, and examples of segmentation results overlaid with OCT images for two grading categories. Fig 2 shows a table of the success rate for each segmented surface. The success rate of the posterior cornea segmentation is highest at 98%. The success rate for all surfaces in the same image is 76%.
Conclusions :
We prototyped a deep learning solution for AC segmentation using a limited number of subjects. The preliminary performance of the algorithm shows that deep learning-based AC segmentation is feasible for generating acceptable AC measurement for disease management.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.