A Deep Learning Model to Improve Pathologist Interpretation of Donor Kidney Biopsies

项目总结/摘要 每年死于肾脏疾病的人比死于乳腺癌或前列腺癌的人多。肾 移植是挽救生命的，但受到器官捐献者短缺和不可接受的 供体器官丢弃率高。决定使用或丢弃捐赠者的肾脏在很大程度上取决于 病理学家对关键显微镜检查结果进行手动定量。这一点的一个主要局限是， 显微镜检查是人类在解释发现时的可变性和低效率， 导致潜在健康的器官被认为不适合移植或潜在地 受损器官被不当移植我们的团队开发了第一个 一种可自动量化肾小球硬化整体百分比的学习模型 供肾冰冻切片楔形活检切片图像。这种创新的方法具有 通过提高病理学家的效率来改变供体肾活检评价的潜力， 准确性和精确性最终导致优化的供体器官利用， 医疗保健成本和改善患者的结果。这个项目的目标是建立我们的 深度学习自动化定量评估作为供体肾脏的标准实践 移植前评估。这将通过组建一个肾脏专家团队来实现 病理学家和专门从事机器学习的计算机科学家。该提案将 评估计算机化方法的准确性和精确性，以量化全球 肾小球硬化症，并将这些结果与当前的护理标准病理学家进行比较 评价部署深度学习模型来分析整个幻灯片图像的可行性 云计算也将受到考验。该STTR的最终产品将是一个基于网络的平台 安全地部署深度学习图像分析作为一种工具，以帮助病理学家进行供体 肾活检评价。

PROJECT SUMMARY/ABSTRACT More people die every year from kidney disease than breast or prostate cancer. Kidney transplantation is life-saving but is limited by a shortage of organ donors and an unacceptably high donor organ discard rate. The decision to use or discard a donor kidney relies heavily on manual quantitation of key microscopic findings by pathologists. A major limitation of this microscopic examination is human variability and inefficiency in interpreting the findings, resulting in potentially healthy organs being deemed unsuitable for transplantation or potentially damaged organs being transplanted inappropriately. Our team developed the first Deep Learning model capable of automatically quantifying percent global glomerulosclerosis in whole slide images of donor kidney frozen section wedge biopsies. This innovative approach has the potential to transform donor kidney biopsy evaluation by improving pathologist efficiency, accuracy, and precision ultimately resulting in optimized donor organ utilization, diminished health care costs, and improved patient outcomes. The goal of this project is to establish our Deep Learning automated quantitative evaluation as the standard practice of donor kidney evaluation prior to transplantation. This will be achieved by assembling a team of expert kidney pathologists and computer scientists specializing in machine learning. The proposal will evaluate the accuracy and precision of the computerized approach to quantifying percent global glomerulosclerosis and compare these results with current standard of care pathologist evaluation. The feasibility of deploying the Deep Learning model to analyze whole slide images on the cloud will also be examined. The end product of this STTR will be a web-based platform to securely deploy Deep Learning image analysis as a tool to assist pathologists with donor kidney biopsy evaluation.