Developing a diagnosis model for dry eye disease in dogs using object detection

Abstract

The purpose of this study was to develop an object detection method for the diagnosis of dry eye disease (DED) in dogs. To this end, a methodology was designed to evaluate ocular surface video images using the YOLOv5 model, which is an object detection algorithm that has been widely used because of its simple network structure and fast detection speed. Because the cornea is a transparent organ, an illuminator plate with grid squares was used to provide grid lines, which were analyzed as the reflected straight lines of the light source representing the precorneal tear film (PTF) stability. The original video consisted of the number of 12 normal images(normal, [Formula: see text] = 17) and the number of 15 abnormal images(abnormal, [Formula: see text] = 17), converted to JPEG images for labeling, learning, and model validation. The labeled image data were divided into a training image data set (normal, [Formula: see text] = 15,276; abnormal, [Formula: see text] = 26,196) to a validation image data set (normal, [Formula: see text] = 6546; abnormal, [Formula: see text] = 11,228). As a result of the experiment, the mean average precision ([Formula: see text]) achieved 0.995. This study proposes a method to effectively determine ocular surface status in dogs by using YOLOv5 and concludes that an object detection model can be used in the veterinary field.