A Second Look at DREAM: Towards a New Paradigm in Meibomian Gland Evaluation Using Artificial Intelligence

项目摘要 干眼症（DE）是一种非常普遍的疾病，对个人和社会具有重大影响， 仍然无法轻易诊断和治疗。DE最常见的原因被认为是睑板腺炎 腺功能障碍（MGD）。上眼睑和下眼睑的睑板腺分泌脂质， 薄膜覆盖含水眼泪，并抑制其蒸发。在MGD中，人们认为， 分泌高质量的泪液脂质，导致泪膜不稳定、蒸发和DE症状。腺 MGD中发生的变化还不清楚，我们也不能确定MGD的哪些方面构成了 泪膜不稳定和DE的最大风险。 干眼评估和管理（DREAM）研究是一项ω 3脂肪酸的临床试验， 补充治疗DE。在整个试验过程中，一个大型的睑板造影图像数据库- 红外图像的外翻眼睑，揭示了睑板腺-是汇编和分析使用 新的计划，以表征13个不同方面的腺体通过目视检查，并分析其 与分泌脂质的临床评估质量的关系。这个过程是艰苦的，时间 消费，固有地受到人类偏见的影响，并提供了关于两者之间联系的新信息。 睑板腺特征和DE体征和症状。 人工智能（AI）的最新进展使我们能够训练监督机器学习 在睑板造影图像上的算法，以自动检测和量化睑板造影图像的详细形态特征。 单个腺体。这些详细的形态学特征潜在地包含了大量关于 健康和功能的睑板腺，并可以提供有价值的信息的机制， MGD及其临床意义的背后。另一种新兴的用于医学成像的人工智能技术- 无监督的判别特征学习-减轻了人类的偏见，并可能发现 睑板造影图像中先前未识别的性质，以及与关键临床终点的可能联系，如 泪膜不稳定和DE症状。 在这个项目中，我们建议利用这种新的人工智能技术来重新分析DREAM研究临床 数据库的睑板造影图像，以显着扩大他们的初步研究结果。具体来说，我们将采用 无监督判别特征学习，以减轻睑板造影分析中的人类偏见，发现 以前未认识到的特征的睑板腺，并分析这些功能之间的联系 和MGD、泪膜不稳定性以及指示DE的临床体征和症状。

Project Summary Dry eye (DE) is a highly prevalent condition with significant impacts on individuals and society that continues to evade easy diagnosis and treatment. The most common cause of DE is thought to be Meibomian gland dysfunction (MGD). The Meibomian glands in the upper and lower eyelids secrete lipids that form a thin film covering the aqueous tears and inhibit their evaporation. In MGD, it is thought that inadequate and/or poor quality tear lipids are secreted, leading to tear film instability, evaporation, and symptoms of DE. The glandular changes that occur in MGD are not well understood, nor are we able to identify which aspects of MGD pose the greatest risk for tear film instability and DE. The Dry Eye Assessment and Management (DREAM) Study was a clinical trial of ω3 fatty acid supplements for the treatment of DE. Over the course of the trial a large database of meibography images – infrared images of the everted eyelids that reveal the Meibomian glands – was compiled and analyzed using a novel scheme to characterize 13 different aspects of the glands by visual inspection and analyze their relationships to the clinically assessed quality of the secreted lipids. The process was arduous and time consuming, inherently subject to human bias, and provided little new information on the links between Meibomian gland characteristics and DE signs and symptoms. Recent advances in artificial intelligence (AI) have allowed us to train supervised machine learning algorithms on meibography images to automatically detect and quantify detailed morphological features of the individual glands. These detailed morphological features potentially contain a wealth of information about the health and functioning of the Meibomian glands, and could provide valuable information on the mechanisms behind MGD and its clinical implications. A further emerging AI technology for use in medical imaging – unsupervised discriminative feature learning – mitigates the human bias, and could potentially discover previously unidentified properties in meibography images, and possible links to crucial clinical endpoints like tear film instability and DE symptoms. In this project, we propose to utilize this new AI technology to re-analyze the DREAM Study clinical database of meibography images to dramatically extend their initial findings. Specifically, we will employ unsupervised discriminative feature learning to mitigate the human bias in meibography analysis, discover previously unrecognized features of the Meibomian glands, and to analyze the links between these features and MGD, tear film instability, and the clinical signs and symptoms indicative of DE.