Automatic quantification of ocular dryness by artificial intelligence in the context of Sjögren’s syndrome

Sjögren’s syndrome is an immune system disorder with two common symptoms, dry eyes and a dry mouth. The discomfort of dry eye symptoms affects daily lives, results in 30% activity impairment and affects 95% of Sjögren patients [1]. Dry eye disease (DED) is also an independent multifactorial disorder with a prevalence of up to 50% [2]. The ocular surface inflammation causes discomfort, fatigue and overall, a lower quality of life [2, 3]. Traditional therapies help manage the symptoms and avoid permanent damage. Hence, it is pivotal to grade and follow the development of DED. A common drawback in existingmethods that diagnose and quantify DED is reproducibility, invasivity and inaccuracy.We reviewed classical methods and those that incorporate automation to measure the extent of DED [4]. The study showed that DED has yet to benefit from what Artificial Intelligence (AI) has to offer. Using slit-lamp examinationsof the ocular surface we aimed to improve the quantification of the Oxford score [5]. Our proposed method uses unsupervised learning to register frames from the examinations to a common coordinate system. By learning the camera motion and depth simultaneously we are able to track the ocular surface in 3-D, compensate for eye motion and visualise the full eye. The lightsource attached to the camera is a challenge and a disturbance when learning egomotion. This was solved through semantic segmentation and adding a new supervision signal: semantic reconstruction loss. We also used the advantage of estimatingthe shape of the eye as prior knowledge we could include as a constraint. This was implemented through a shape fitting loss; the shapes being two spheres intersecting each other. Our registration showed quantitative and qualitative improvement with each contribution. We also calculated the inter-rater reliability of the punctate dots (damaged areas) annotations. Our method came closest to what can be considered human error. The proposed registration method was also used for a pre-processing task, frame selection. Once applied to automated Oxford score classification, our method improved the results as well. The improvement validates that the strong color/illumination variances present in the examinations are a disturbance for any deep learning task. We overcame this in both tasks via our contributions and proposed method.