RetinaScan: AI-enabled automated image assessment system for diabetic retinopathy screening

RetinaScan：用于糖尿病视网膜病变筛查的人工智能自动图像评估系统

• 批准号: 104184
• 金额: $ 100.86万
• 依托单位: RETINASCAN LIMITED
• 依托单位国家: 英国
• 项目类别: Collaborative R&D
• 财政年份: 2018
• 资助国家: 英国
• 起止时间: 2018 至 无数据
• 项目状态: 已结题
• 来源: https://gtr.ukri.org/projects?ref=104184
• 关键词: RetinaScan; AI; enabled; automated; image

"Diabetes mellitus (DM) is a global healthcare problem. In 2014 there were 422m diabetics, forecast to rise to 642m by 2040 \[WHO\]. Diabetic Retinopathy (DR) is a common complication (\>50% of sufferers) caused by physiological changes in the retina. It is a major cause of blindness (\>7% UK blindness) but easily ameliorated through laser or drug treatment.Annual routine screening enables DR to be captured and treated early. Images of the retina are taken using readily available cameras for qualified people to review for symptomatic features. Few countries have managed to run a diabetic eye screening programme (DESP), the UK being one. Whilst highly effective, current DESPs are:* labour intensive, requiring manual grading of images by up to 3 specialists* slow, with a targeted 6-week turnaround, impacting on patient retention* expensive, generating an annual NHS screening cost of \>£100mAutomated retinal image analysis systems (ARIAS) utilise image analysis algorithms to detect disease features. ARIAS have shown potential to transform DESP delivery with speed, non-scaling cost of operation and significant cost savings. However, achieving accuracy at the level required to provide an effective replacement of level 1 human grading has not yet been realised.RetinaScan meets this challenge through an innovative ARIAS solution: an advanced algorithm methodology design by experts in diabetology associated with the University of Oxford; with novel AI based imaging analysis systems (convolution neural networks - CNNs) developed at the University of Surrey. Led by Retinopathy Answer Ltd (RAL), a proof of concept prototype has been devised and validated. RALs Augmented-CNN design brings together deep understanding of imaging data and methods to automate consistently qualified human levels of performance.Advancing on this prototype, RetinaScan will: i) further develop the system architecture for end-user scenarios; ii) advance retina scan datasets for system training and testing; iii) develop regional and eye level image processing engines for accurate disease grading; and iv) develop and demonstrate a complete web-based system prototype through user trials.The key outcomes of the research will be: a fully capable prototype suite of trained Augmented-CNN technology; ii) prototype validation via user trials; and a detailed business plan for commercialisation as a cloud-based service for markets globally.The potential addressable global market for ARIAS is estimated at \>£2.98 billion. The partnership targets ~£14.98 million business growth within a 5-year period (~£27.89m cumulative sales), creating \>35 new jobs and generating a \>30-fold ROI."

“糖尿病（DM）是一个全球性的健康问题。2014年，全球有4.22亿糖尿病患者，预计到2040年将增至6.42亿[WHO]。糖尿病视网膜病变（Diabetic Retinopathy，DR）是一种常见的并发症，约占糖尿病患者的50%以上。它是致盲的主要原因之一（>7%的英国人失明），但通过激光或药物治疗很容易得到改善。视网膜的图像是使用现成的相机为合格的人审查的症状特征。很少有国家成功地运行了糖尿病眼睛筛查计划（DESP），英国就是其中之一。虽然非常有效，但目前的DESP是：* 劳动密集型的，需要多达3名专家对图像进行手动分级 * 缓慢，目标是6周的周转时间，影响患者保留 * 昂贵，每年产生的NHS筛查成本> 1亿英镑自动视网膜图像分析系统（阿里亚斯）利用图像分析算法来检测疾病特征。阿里亚斯已经显示出以快速、非规模化运营成本和显著的成本节约来改变DESP交付的潜力。RetinaScan通过创新的阿里亚斯解决方案来应对这一挑战：由牛津大学糖尿病专家设计的先进算法方法;由萨里大学开发的基于人工智能的成像分析系统（卷积神经网络-CNN）。在Retinopathy Answer Ltd（RAL）的领导下，设计并验证了概念原型的证明。RALs Augmented-CNN设计将对成像数据和方法的深刻理解结合在一起，以自动化一致合格的人类性能水平。在此原型的基础上，RetinaScan将：i）进一步开发最终用户场景的系统架构; ii）推进视网膜扫描数据集的系统培训和测试; iii）开发区域和眼睛级别的图像处理引擎，以准确进行疾病分级; iii）开发用于诊断的图像处理引擎。以及iv）通过用户试验开发和演示一个完整的基于网络的系统原型。研究的主要成果将是：一个完全有能力的训练增强CNN技术原型套件; ii）通过用户试验验证原型;以及一份详细的商业计划，作为面向全球市场的基于云的服务进行商业化。阿里亚斯的潜在可寻址全球市场估计为\超过29.8亿英镑。该合作伙伴关系的目标是在5年内实现约1498万英镑的业务增长（累计销售额约为2789万英镑），创造35个新工作岗位，并产生30倍的投资回报率。"