Abstract
Purpose :
This study describes the application of deep learning methodology to identify the pupil, limbus, and eyelid boundaries in images from a variety of commercially available topographers with a single unified software approach.
Methods :
604 de-identified unique grayscale infrared iris images were obtained from a variety of commercially available topographers. All images had the boundaries delineating the pupil and visible iris annotated by hand using software tools. A custom pupil warping method was utilized to produce augmented copies of each image with the iris “stretched” to varying pupil shapes and sizes. A deep convolutional neural network based on the U-Net architecture was trained to label pixels from these images as belonging to the pupil, iris or neither. An active contours algorithm was developed to post-process the U-Net output into geometric curves and infer eyelid interference based on the shape of the iris contour. Performance was evaluated through two-fold cross-validation based on geometric center accuracy and dice coefficient (F1 score) of the final curve-bounded regions.
Results :
The pupil and limbus were correctly identified in 603 of 604 unique eyes for a success rate of 99.8%. The pupil center error – defined as the Euclidean distance between the center-of-mass of the hand-labeled pupil region and the center of the best-fit ellipse to the identified pupil curve – was below 5 pixels in 98.7% of cases, below 10 pixels in 99.5% of cases, and below 15 pixels in 99.7% of cases (pixel scaling varies from approximately 0.03 to 0.045 mm per pixel). The similarly defined limbus center error was below 5 pixels in 69.2% of cases, below 10 pixels in 95.0% of cases, and below 15 pixels in 99.8% of cases. Dice coefficients measured above 0.95 in 95.6% of cases for the pupil, 98.2% for the composite iris region (including the pupil area), and 92.0% for the exclusive iris region (excluding the pupil area).
Conclusions :
Modern deep learning technology is capable of segmenting iris images from a wide variety of diagnostic devices under a unified software approach, with a standardized image resizing scheme serving as the only device-specific operation. This technological advance has significant implications for device interoperability, as it can enable device software to handle images from multiple additional input devices with little or no modification.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.