SBIR Phase I: Novel machine learning framework for the classification of non-mydriatic retinal images

这个小企业创新研究（SBIR）第一阶段项目的更广泛的影响/商业潜力是一种基于人工智能（AI）的方法来筛查糖尿病视网膜病变（DR），这是美国视力障碍和失明的主要原因。DR影响美国近400万人，每年直接成本近5亿美元。如果早期诊断，临床治疗和生活方式的改变可以阻止疾病的进展，防止失明。然而，视网膜检查目前需要昂贵的设备和侵入性的眼睛扩张，这限制了眼科或验光实践的筛查，导致病情的诊断不足，特别是在服务不足的人群中。该项目提出了一种新的相机和机器学习方法的系统，使临床医生能够识别DR和其他视网膜疾病。这个小企业创新研究（SBIR）第一阶段项目旨在探索开发一种新型软件驱动的非散瞳眼底相机的可行性，可以识别DR。拟议的创新基于：1）便携式相机，其使用人眼不可见的近红外（NIR）光来照射视网膜并获取眼底图像，使得非专业人员能够使用该设备; 2）一种基于迁移学习的新框架，它用有限的训练数据（100张图像）训练神经网络。在这个项目中，一个原型系统将收集近红外视网膜图像，目标是开发一种能够处理这些图像的AI分类算法。与此同时，将开发一种新的图像处理算法，以提高近红外图像的分辨率，基于对比度归一化方法和降噪技术。该奖项反映了NSF的法定使命，并已被认为值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估的支持。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project is an artificial intelligence (AI)-based method to screen diabetic retinopathy (DR), the leading cause of vision impairment and blindness in the US. DR affects almost 4 million people in the US and is associated with direct annual costs of almost $500 M. If diagnosed early, clinical treatment and lifestyle changes can halt the progression of the disease, preventing blindness. However, retinal exams currently require expensive equipment and invasive eye dilation that restrict screenings to ophthalmology or optometry practices, leading to the under-diagnosis of the condition, particularly in underserved populations. This project advances a system with a new camera and a machine learning approach to enable recognition of DR and other retinal disorders by clinicians.This Small Business Innovation Research (SBIR) Phase I project seeks to explore the feasibility of developing a novel software-enabled non-mydriatic fundus camera that can identifiy DR. The proposed innovation is based on: 1) a portable camera that uses near-infrared (NIR) light, invisible to the human eye, to illuminate the retina and acquire fundus images, enabling the use of the device by non-specialists; 2) a novel framework based on transfer learning, which trains Neural Networks with a limited amount of training data (100 images). In this project, a prototype system will collect NIR retinal images, with the goal of developing an AI classification algorithm capable of processing these images. In parallel, a new image processing algorithm will be developed to improve the resolution of the NIR images, based on contrast normalization methods and noise reduction techniques.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.