Abstract
Purpose :
Blinking is a recognized indicator of ocular surface diseases. However, detailed analysis of blink dynamics is complicated. In this study, we endeavored to investigate the blink patterns in dogs by employing deep learning techniques on smart phone-captured videos.
Methods :
Ten eyes from five healthy beagle dogs were the focus of our study. Utilizing DeepLabCut™, we marked five key points (upper and lower eyelids of both eyes, along with a base point) by continuously recording the blinking eyes for one minute using an iPhone XR (4K, 60fps). The labeled frames underwent training with ResNet-50 for 80,000 iterations. Subsequently, each one-minute video was meticulously analyzed, and the coordinates of each tracking point were documented.
Results :
DeepLabCutTM facilitated the calculation of various parameters such as eyelid cleft width, ECR (eyelid closure rate), UECR (upper eyelid closure rate), and LECR (lower eyelid closure rate) for both left and right eyelids. Blink categorization was based on ECR percentages, classifying blinks as minimal (8%-50%), moderate (50%-88%), perfect (88%-95%), and squeeze (>95%). Comparative analysis of blink classifications revealed distinctions in ECR, blink duration, and orbicularis oculus muscle contraction strength between complete (squeeze) and incomplete blinks. Approximately 88% of blinks were categorized as incomplete (minimal) and moderate (moderate), showcasing a greater range of motion of the lower eyelid in canines. Unilateral blinks occurred in about 20% of cases, and even in bilateral blinks, left-right differences in ECR were observed.
Conclusions :
Our study demonstrates the feasibility of employing deep learning for canine eyelid and blink tracking, providing valuable insights into detailed blink features in dogs. This novel method may be useful not only for blink dynamics but also for understanding the pathogenesis of ocular surface diseases.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.