Deep learning to quantify glaucomatous damage on fundus photographs for teleophthalmology

深度学习量化眼底照片上的青光眼损伤，用于远程眼科

• 批准号: 10415277
• 负责人: Atalie C Thompson
• 金额: $ 17.17万
• 依托单位: WAKE FOREST UNIVERSITY HEALTH SCIENCES
• 依托单位国家: 美国
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-06-02 至 2025-02-28
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10415277
• 关键词: Address; Adult; Advisory Committees; Agreement; Algorithms; Artificial Intelligence; Assisted Living Facilities; Biomedical Engineering; Biometry; Blindness; Clinic; Clinical; Clinical Research; Consumption; Data; Dependence; Detection; Development; Diabetic Retinopathy; Diagnosis; Diagnostic; Discrimination; Disease; Early Diagnosis; Effectiveness; Environment; Epidemiology; Evaluation; Eye; Eye diseases; Fellowship; Frequencies; Fundus; Fundus photography; Glaucoma; Goals; Gold; Grant; Human; Image; Imaging technology; Improve Access; Individual; Label; Machine Learning; Manuals; Masks; Master' s Degree; Measurement; Medical; Medicine; Mentored Patient-Oriented Research Career Development Award; Mentors; Mentorship; Nature; Optic Disk; Optical Coherence Tomography; Output; Patients; Perimetry; Primary Health Care; Public Health; Reference Standards; Reproducibility; Research; Research Priority; Research Proposals; Resources; Scientist; Screening procedure; Sensitivity and Specificity; Severity of illness; Specialist; Suspect Glaucomas; Technology; Testing; Thick; Time; Training; Validation; Visual Fields; Visual impairment; Width; Work; algorithm training; career; carina; cohort; cost; cost effective; deep learning; deep learning algorithm; deep neural network; demographics; experience; eye center; health care disparity; high risk; improved; innovation; intelligent algorithm; interest; learning network; neural network; novel; novel diagnostics; population based; programs; prospective; public health intervention; responsible research conduct; retinal nerve fiber layer; screening; teleophthalmology; tool

PROJECT SUMMARY/ABSTRACT Candidate: Atalie Carina Thompson, MD, MPH is a current glaucoma fellow and Heed fellow with a long-term career goal of becoming an independent clinician-scientist and leader in the field of glaucoma and public health. She has a long-standing interest in addressing healthcare disparities in medicine, and in improving the diagnosis of glaucoma and other ophthalmic diseases through imaging technology. While obtaining a medical degree at Stanford, she received a fellowship to complete a master’s degree in public health with additional higher-level coursework in biostatistics and epidemiology. Her immediate goal in this proposal is to refine and validate a deep learning (DL) algorithm capable of quantifying neuroretinal damage on optic disc photographs and then to apply it in a pilot teleophthalmology program. With a K23 Mentored Patient-Oriented Research Career Development Award, she will acquire additional didactic training and mentored research experience in glaucoma imaging, machine learning, biostatistics, clinical research, and the responsible conduct of research. Environment: The mentorship and expertise of the advisory committee, the extensive resources at the Duke Eye Center and Departments of Biostatistics and Biomedical Engineering, and the significant institutional commitment will provide her with the support needed to transition successfully into an independent clinician-scientist. Research: This proposal will test the hypothesis that a DL algorithm trained with SDOCT detects glaucoma on optic disc photographs with greater accuracy than human graders. In Specific Aim 1, a DL algorithm that quantifies neuroretinal damage on optic disc photographs will be refined. The main hypothesis is that the quantitative output provided by the DL algorithm will allow accurate discrimination of eyes at different stages of the disease according to standard automated perimetry, and will generate cut-offs suitable for use in a screening setting. In Specific Aim 2, the short-term repeatability and reproducibility of the DL algorithm in optic disc photographs acquired over a time period of several weeks will be determined. The hypothesis is that the test-retest variability of the predictions from the DL algorithm will be similar to the original measurements acquired by SDOCT. In Specific Aim 3, the DL algorithm will be applied to optic disc photographs obtained during a pilot screening teleophthalmology program in primary care clinics and assisted living facilities. The hypothesis is that the DL algorithm will be more accurate than human graders when a full ophthalmic examination is used as the gold standard. This work will constitute the basis of an R01 grant and will advance our understanding of the application of deep learning algorithms in glaucoma and teleophthalmology.

项目摘要/摘要 候选人：Atalie Carina Thompson，医学博士，公共卫生硕士，目前是青光眼研究员，也是长期 职业目标是成为一名独立的临床医生--青光眼和公共卫生领域的科学家和领导者。 她长期以来一直对解决医疗保健方面的医疗差距和改善诊断感兴趣 通过成像技术诊断青光眼和其他眼科疾病。在获得医学学位的同时 在斯坦福大学，她获得了奖学金，完成了公共卫生硕士学位，并获得了更高级别的额外学位 生物统计学和流行病学的课程。她在这项提议中的直接目标是提炼和验证一个深层次的 能够对视盘照片上的神经视网膜损伤进行量化的学习(DL)算法 它参与了一项远程眼科试验项目。以K23为导师，以患者为导向的研究生涯发展 获奖后，她将获得额外的教学培训和青光眼成像方面的指导研究经验， 机器学习、生物统计学、临床研究和负责任的研究行为。环境： 咨询委员会的指导和专业知识，杜克眼科中心和 生物统计和生物医学工程系，以及重大的机构承诺将 为她提供成功转型为独立临床医生兼科学家所需的支持。研究： 这项提议将检验这样一个假设，即用SDOCT训练的DL算法可以检测到视盘上的青光眼 照片比人类评分员的准确度更高。在特定目标1中，一个量化的DL算法 视盘照片上的神经视网膜损伤将被细化。主要的假设是，数量产出 由DL算法提供，将允许在疾病的不同阶段准确区分眼睛 根据标准的自动视野检查，并将生成适合在筛查设置中使用的截止值。在……里面 特定目标2：DL算法在视盘照片中的短期重复性和再现性 在几个星期的时间内获得的将被确定。假设是重测的变异性 来自DL算法的预测结果将与SDOCT获取的原始测量值相似。在……里面 具体目标3，DL算法将应用于在试点筛选期间获得的视盘照片 初级保健诊所和辅助生活设施的远程眼科计划。假设是DL值 当使用全面眼科检查作为黄金时，算法将比人类评分器更准确 标准的。这项工作将构成R01赠款的基础，并将促进我们对应用程序的理解 深度学习算法在青光眼和远程眼科中的应用。