A Comparison of YOLOv8 and DEtection TRansformer in the Context of Tear Meniscus Height Calculation Based on Ocular Surface Videos

This study is the first investigation related to the application of advanced object detection algorithms, specifically You Only Look Once version 8 (YOLOv8) and DEtection TRansformer (DETR), in the realm of TMH detection. The research leverages a dataset comprising 152 ocular surface videos captured by the Keratograph 5M, undertaking a thorough comparison of architectural distinctions, advantages, and drawbacks between YOLOv8 and DETR. Performance metrics, encompassing mean Average Precision (mAP), specificity, recall, F1-score, and Frames Per Second (FPS), were meticulously evaluated, with an input size set at 640x640 pixels. The inquiry delves into the potential benefits of implementing AI-based methods for TMH detection, emphasizing their applications in ophthalmology and the management of dry eye conditions. The utilization of YOLOv8 and DETR algorithms facilitates automated analysis of anterior segment videos, markedly diminishing the manual workload for clinicians. Furthermore, these methodologies yield objective and consistent TMH measurements, thereby refining the precision of assessments and bolstering the reliability of dry eye diagnosis. Additionally, AI-based methods enable early detection of TMH alterations and timely interventions in cases of dry eye syndrome.