Exploring Clinically-relevant Image Retrieval for Diabetic Retinopathy Diagnosis

项目摘要 所有糖尿病患者都有患糖尿病视网膜病变的风险，这是一种威胁视力的并发症。 尽管多年来糖尿病护理取得了进展，但糖尿病视网膜病变仍然是一种潜在的破坏性疾病， 并发症早期发现和及时干预或治疗可减少因 糖尿病性视网膜病变近年来，基于数字视网膜成像的诊断已成为一种替代方法 传统的面对面评估。数字图像自动分析的潜在好处 糖尿病性视网膜病变已经在现有的研究中显示。然而，目前没有基于计算机的系统， 可以达到人类专家的性能水平。 这一建议采取了一个新的角度，在发展一个计算机辅助系统，以改善诊断的 糖尿病视网膜病变，通过探索新的计算方法，从 对于给定的新图像，具有先前诊断信息的存档数据库。图像被视为 如果它们包含具有相似严重程度的相同类型病变，则具有临床相关性。研究和 基于内容的视网膜图像搜索/检索的发展仍处于起步阶段，仅取得有限的成功， 这主要是由于将专家知识显式编码到计算算法中的挑战。处理 面对挑战，这个研究项目采取了一种截然不同的方法， 方法，其中标记的图像集用于训练用于分析其他新图像的计算机算法， 训练的重点是临床相关性的相似性而不是图像特征。培训已启用 部分原因是研究人员对基于计算机的病变模拟的现有研究。 本研究的一个具体目标是建立一个基于内容的图像检索系统， 临床医生即时参考与诊断下的图像临床相关的存档图像。 这是一种创新的方式，可以利用隐藏在先前诊断的图书馆中的大量专家知识。 糖尿病视网膜病变的数字图像，以提高临床医生的诊断性能。另一个具体 目的是建立一个糖尿病视网膜病变图像信息管理系统， 在实际临床环境中展开回收系统。除了检索系统外， 研究成果还包括糖尿病视网膜病变图像的自动评估算法， 与现有方法相比，潜在地改进了性能。特别是，建议的设计 工作允许根据医生的具体需要对所得到的系统进行不同的配置。

Project Summary All people with diabetes have the risk of developing diabetic retinopathy, a vision-threatening complication. Despite advances in diabetes care over the years, diabetic retinopathy remains a potentially devastating complication. Early detection and timely intervention or treatment can reduce the incidence of blindness due to diabetic retinopathy. Recent years, diagnosis based on digital retinal imaging has become an alternative to traditional face-to-face evaluation. The potential benefits of automated analysis of digital images of diabetic retinopathy have been shown in existing studies. However, no current computer-based systems can achieve the same level of performance of human experts. This proposal takes a new perspective in developing a computer-aided system for improved diagnosis of diabetic retinopathy, by exploring novel computational methods for retrieving clinically-relevant images from archived database with prior diagnosis information, for a given novel image. Images are considered as being clinically relevant if they contain the same types of lesions with similar severity levels. Research and development on content-based retinal image search/retrieval is still in its infancy, with only limited success, largely due to the challenge of explicitly coding expert-knowledge into a computational algorithm. To deal with the challenge, this research project takes a distinctly different approach engaging a machine-learning approach, where a labeled image set is used to train a computer algorithm for analyzing other new images, with the focus of training on similarity in clinical relevance instead of image features. The training is enabled in part by the investigators' existing research on computer-based lesion simulation. One specific aim of the research is to build a content-based image retrieval system that can provide a clinician with instant reference to archival images that are clinically relevant to the image under diagnosis. This is an innovative way of exploiting vast expert knowledge hidden in libraries of previously-diagnosed digital images of diabetic retinopathy for a clinician's improved performance in diagnosis. Another specific aim is to build an image information management system for diabetic retinopathy that supports the deployment of the retrieval system in a realistic clinical setting. In addition to the retrieval system, the direct outcome of the research also includes automated evaluation algorithms for diabetic retinopathy images with potentially improved performance compared with existing methods. In particular, the design of the proposed work allows different configurations of the resultant system according to the specific needs of a physician.