Abstract
Purpose :
Optical coherence tomography images of the cornea may include non-corneal and non-physiologic features and artifacts such as iris, lens, and saturation artifacts. The robust segmentation of the cornea from such OCT images despite these challenges can enable objective corneal evaluation for both research and clinical applications. Deep learning techniques have been used to segment ophthalmic structures without ad hoc, rule-based approaches. This study aims to develop and evaluate a pipeline for automated segmentation of the cornea from artifact-containing OCTs leveraging deep learning with a convolutional neural network.
Methods :
Three hundred mouse anterior segment OCT radial scans (Bioptigen) were obtained from 4 volumes. Images contained artifacts including reflex saturation and the presence of lens, iris, tear lake, eyelids, and eyelashes. All images were first manually labeled to generate ground truth semantic segmentations and ground truth contours for the anterior and posterior corneal surfaces. Fifty images from a single volume were utilized to train a deep convolutional neural network implemented in Python with Keras and Tensorflow. Network semantic segmentations of the remaining 3 volumes (none involving the same eye as the training volume) were post-processed using morphological techniques to generate predicted contours of the anterior and posterior corneal surfaces. Metrics were computed comparing the ground truth and machine-generated segmentations and contour predictions, including pixel error, rand error, warping error, and intersection over union.
Results :
On the 200 images of the testing set the network achieved an intersection over union score of 0.74, indicating segmentation of reasonable quality despite the presence of the aforementioned artifacts. Rand error was 0.086, pixel error was 0.14, and warping error of the anterior and posterior contours was 0.0017.
Conclusions :
Corneal segmentation in the presence of non-corneal or non-physiologic structures presents a greater challenge than corneal segmentation in the isolated normal cornea. The pipeline described here utilizes deep learning for robust semantic segmentation of the cornea despite the presence of non-corneal structures such as lens, iris, tear lake, eyelids, and eyelashes, and may be applicable to a range of pathologic states of the cornea.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.