Dry eye (DE) is a very common multifactorial disorder characterized by loss of tear film homeostasis that may result in discomfort and visual disturbances caused by tear film instability, increased osmolarity with ocular surface inflammation and damage, and neuroparesthesia [1]. In clinical practice, most diagnostic tests are based on measuring tear film using minimally invasive techniques such as staining by fluorescein and Schirmer's test. Recent advances in new technologies have allowed non-invasive evaluation of the quantity and quality of tear film thereby addressing many of the limitations inherent to invasive tests [2].

The keratograph 5 M (Oculus Gmbh, Wetzlar, Germany) is a Placido disc-based corneal topographer with high-definition color camera optimized for external imaging that provides a multifaceted non-invasive assessment of the tear film. The machine features a high-resolution color camera and integrated magnification changer allowing ocular surface photography and video capture under white, blue, or infra-red illumination systems. This allows objective and subjective assessment of tear film volume (Tear Meniscus Height - TMH), tear film stability (Non-Invasive Keratograph Break-Up Time – NIKBUT), Meibomian gland observation (Meiboscan) and, Tear Film Lipid Layer Examination using the principle of white light interferometry [3]. This non-contact keratograph provides simple non-invasive screening tests for DE with acceptable sensitivity, specificity, and repeatability [4].

Deep learning has greatly improved the diagnostic accuracy for several ocular diseases [5,6]. Recently, deep learning applied to video sequences for rapid screening of patients undergoing ocular ultrasonography has provided high diagnostic accuracy [7].

In our study, we assessed the feasibility of employing deep learning models based on single video frames classification for automated DE diagnosis in patients undergoing video recording by the Keratograph 5 M.