Automated retinopathy of prematurity classification using machine learning

描述(由申请人提供)：该项目的目标是开发一种基于网络的半自动系统，利用从NIH资助的先前研究中收集的现有视网膜图像集，识别严重的早产儿视网膜病变(ROP)。如果及早诊断，ROP是可以治疗的，但它仍然是全世界儿童失明的主要原因。ROP的眼底镜检查结果的诊断和记录是主观和定性的，研究发现，即使向专家展示了完全相同的临床数据，诊断结果往往也存在显著差异。基于计算机的图像分析和将机器学习技术应用于特征提取和图像分类，有可能解决许多这些限制。图像处理的最新进展已经导致了跟踪血管状结构的复杂技术。此外，机器学习技术将使我们能够利用这些现有的注释图像数据库来提高我们的血管分割和疾病分类算法的性能。我们的总体假设是，视网膜血管特征可以被量化，并用于辅助临床医生诊断ROP。这些假设将通过两个特定的目标进行测试：(1)开发和评估分割视网膜血管的半自动算法，并生成一组基于视网膜血管的特征。(2)开发与专家意见最相关的计算机决策支持算法。总体而言，该项目将建立在先前研究基础设施的基础上，为提高临床ROP诊断的准确性和一致性创造潜力，提供现实世界医疗过程中图像分析的基于计算机的决策支持的演示，并刺激未来的研究，以了解与严重ROP相关的血管特征。这个项目将由一个拥有眼科、生物医学信息学、计算机科学、机器学习和图像处理专业知识的多学科调查团队执行。

DESCRIPTION (provided by applicant): The goal of this project is to develop a web-based, semi-automated system for identifying severe retinopathy of prematurity (ROP) with "plus disease," using an existing data set of retinal images collected from previous NIH-funded research studies. ROP is treatable if diagnosed early, yet continues to be a leading cause of childhood blindness throughout the world. Diagnosis and documentation of ophthalmoscopic findings in ROP are subjective and qualitative, and studies have found that there is often significant diagnostic variation, even when experts are shown the exact same clinical data. Computer-based image analysis and the application of machine learning techniques to feature extraction and image classification have potential to address many of these limitations. Recent advances in image processing have had led to sophisticated techniques for tracing vessel-like structures. Additionally, machine-learning techniques will enable us to leverage these existing annotated image databases to improve the performance of our algorithms for vessel segmentation and disease classification. Our overall hypothesis is that retinal vascular features may be quantified and used to assist clinicians in the diagnosis of ROP. These hypotheses will be tested using two Specific Aims: (1) Develop and evaluate semi-automated algorithms to segment retinal vessels and generate a set of retinal vessel-based features. (2) Develop computer-based decision support algorithms that best correlate with expert opinions. Overall, this project will build upon infrastructure developed from previous studies, create potential for improving the accuracy and consistency of clinical ROP diagnosis, provide a demonstration of computer-based decision support from image analysis during real-world medical care, and stimulate future research toward understanding the vascular features associated with severe ROP. This project will be performed by a multi-disciplinary team of investigators with expertise in ophthalmology, biomedical informatics, computer science, machine learning, and image processing.