Advanced Image Analysis Tools for Diabetic Retinopathy Telemedicine Applications

摘要 在这个小企业创新研究（SBIR）项目中，我们提出了EyeArt，一种视网膜 用于自动糖尿病视网膜病变（DR）筛查的图像分析工具， nostic功效凭借其与EyePACS的接口，EyePACS是一个免许可、可扩展的远程医疗平台， 形式，EyeArt将有助于DR筛查的扩展， 糖尿病患者数量与眼科护理数量之间的差距越来越大 提供商 研究表明，拉丁裔人口一般都有遗传倾向， 患上糖尿病他们对糖尿病视网膜病变所致视力丧失的脆弱性进一步增加。 由于缺乏眼科医生，缺乏信息， 安全，缺乏教育。根据卫生服务部（DHS） 在洛杉矶县（LAC），糖尿病患者的情况尤其严峻， 视网膜病变筛查的视网膜检查等待时间超过6-9个月。 这可能导致治疗延迟和不可逆转的视力丧失。到 为了降低糖尿病患者视力丧失的风险，我们建议使用先进的 图像分析算法与现有远程医疗倡议相结合， 更快的筛查，允许眼科医生预约的重新优先次序，并有助于分类 高风险患者。我们的第一阶段原型自动DR筛查工具已经 通过击败当前的学术和商业DR筛查应用显示出巨大的潜力， 在大型公共视网膜数据集上进行研究。展望未来，我们将继续采用我们的方法， 并进一步开发创新的定制算法，用于关键的低级别图像处理， 利用计算机视觉和机器视觉的最新进展， 高级推理步骤的学习区域，以生成临床级DR筛查 工具.我们的病变定位和筛查引擎将在功能上与 EyePACS进一步推动筛查的扩展，特别是在以下方面受益- 资源筛选计划，如LAC-DHS安全网及其大型西班牙裔dia， 贝蒂克人口。

Abstract In this small business innovations research (SBIR) project we present EyeArt, a retinal image analysis tool for automated diabetic retinopathy (DR) screenings with high diag- nostic efficacy. With its interface to EyePACS, a license-free, scalable telemedicine plat- form, EyeArt will aid the expansion of DR screening and help bridge the exponentially growing disparity between the number of diabetic patients and the number of eye-care providers. Research suggests that the Latino population in general are genetically predisposed to develop diabetes. Their vulnerability to vision loss due to diabetic retinopathy is further compounded by factors such as lack of access to ophthalmology clinicians, lack of in- surance, and lack of education. According to the Department of Health Services (DHS) in Los Angeles County (LAC) the situation for diabetics is particularly grim, with current wait times upwards of 6-9 months for retinal examinations for retinopathy screening. This can lead to treatment delays and progression towards irreversible vision loss. To help reduce risk of vision loss in this diabetic population, we propose to use advanced image analysis algorithms in conjunction with existing telemedicine initiatives to enable faster screening, allow reprioritization of ophthalmologist appointments, and aid in triage of high-risk patients. Our phase I prototype automatic DR screening tool has already shown great potential by beating current academic and commercial DR screening ap- proaches on large public retinal datasets. Going forward, we will build on our approach and further develop innovative, customized algorithms for critical low-level image pro- cessing steps, while leveraging on recent advances in computer vision, and machine learning areas for high-level, inference steps to produce a clinical grade DR screening tool. Our lesion localization and screening engine will be functionally integrated with EyePACS to further drive the expansion of screening, particularly benefiting under- resourced screening programs like the LAC-DHS safety net and its large Hispanic dia- betic population.