New Model Of Artificial Intelligence For Theranostic (Diagnostic And Therapeutic) Purposes In Dry Eye

Dry eye Disease (DED) is a multifactorial issue and occurs when one or more of the three layers of a normal tear film including, aqueous fluid, mucus, and oily lipids are disrupted . Depending on the climate, DED can affect up to half of the population of a country and may soon become a worldwide concern in the field of ophthalmology . The prevalence of Dry Eye Disease (DED) has increased in the COVID-19 era, particularly in teenagers. 26-70% of Video display terminal (VDT) users have a high risk of DED due to a low rate of blinking. Our team aim to use Artificial intelligence (AI) for Diagnostic and treatment (Theranostic) of dry eye disease with a simple photo of phone's camera.

For this study, 1010 eye photos from 510 patients has been took with 2 different phones in resolution (minimum 12 MP and maximum 64 MP) from a fixed distance with flashlight on. They were annotated using the Open-CV python library and label-studio application with segmentation of the Iris and Sclera. 20 to 30 dots representing the edge between the eye and the lower eyelid, and three dots representing the perimeter of the iris.

In this study, we used U-Net architecture for semantic segmentation tasks. First, we trained a model used the dice coefficient as a loss function. Then we used the output of this model to crop the bottom part of the iris and lower eyelid. Finally, we used the cropped image to train another segmentation model with U-Net architecture. To find the center of light reflection in the middle of the pupil, we used the first U-Net model's output to find the iris and performed a thresholding function that returns the largest mask of the light reflection. Next, we use this point to draw a vertical line and find the interception point of the vertical line and the edge of the lower eyelid.

After detecting the location of the tear meniscus, with the following steps: 1. Using a thresholding function to find the light reflection contours near the edge of the lower eyelid (there are multiple light reflections, but one of them is the tear meniscus) 2. We exclude the contours with an area under 0.128, heights more than 0.5 or less than 0.05, or contours in which more than half of their area is under the eyelid edge. 3. We select the closest contour to the interception point as the tear meniscus, we find the smallest rectangle surrounding the tear meniscus's contour (Figure 1). Dice coefficient was our criteria for authenticate this model and it reaches 96 %.

Finally, we calculate the Tear Meniscus Height (TMH) with a formula which considers the range of 10.2 to 13 mm for the human iris, we presume the size of the iris to be 11 mm, which results in a measurement error of 1.5 mm, which is acceptable. All the measured photos have been authenticated by an ophthalmologist and this model possesses 98 % of accurate diagnostic of TMH. Now our team is writing an application (can be use on phones) for grading and treating dry eye disease based on TMH and 5 questions about the patient's last week eye condition.