Development and Validation of a Deep Learning system to estimate Interstitial Fibrosis from a kidney ultrasonography image

项目摘要 间质纤维化是肾活检的常见发现，并强烈预测未来肾功能的下降 无论肾脏疾病的潜在病因如何。不幸的是，间质性纤维化很难被捕获， 目前肾功能的临床生物标志物（eGFR和白蛋白尿）。因此，间质纤维化是常见的， 具有重要的预后意义，但临床医生对其存在或严重程度视而不见，除非在罕见的 在进行肾活检的情况下。与此同时，新的药物正在测试，以限制肾脏 间质纤维化，但没有非侵入性方法来评估纤维化随时间的变化。间质 目前，纤维化是通过肾活检的组织病理学检查来估计的，这很少进行。一 评估间质纤维化的非侵入性测试目前不可用。我们令人兴奋的初步数据 证明了使用肾脏的常规超声检查（USG），通过深度学习/人工 智能可以非侵入性地评估间质纤维化的存在和严重程度。的首要目标 本研究旨在进一步开发并在内部和外部验证一种基于深度学习的算法， 相对于肾活检金标准，来自肾脏USG图像的间质纤维化。我们假设 肾USG图像中嵌入的是间质性纤维化相关物，可以通过深度分析提取， 学习和定量分析，以高精度估计间质纤维化，并将改善预测 肾功能的纵向下降。如果是这样的话，考虑到肾脏USG在世界范围内的广泛可用性， 间质性纤维化的非侵入性估计将具有直接的临床意义， 预测以及连续监测治疗反应的间质纤维化的能力。拟议程序 研究将针对三个具体目标：目标1。为了进一步开发和内部验证深度学习- 基于系统的肾USG图像间质纤维化定量。在目标2中，我们将从外部验证 使用USG图像和肾脏活检的独立队列的深度学习模型的性能， 并评估年龄、性别和体型各阶层的表现。最后，在目标3中，我们将确定 基于USG深度学习的间质纤维化评分与肾脏疾病进展相关， 相对于肾活检评估间质纤维化的优势。在完成本计划后， 研究，我们设想开发一个应用程序/插件的超声阅读模块，将促进 深度学习工具的广泛传播，使得基于USG的纤维化评分广泛适用于 治疗临床医生。

PROJECT SUMMARY Interstitial fibrosis is a common finding on kidney biopsy, and strongly predicts future decline in kidney function irrespective of the underlying etiology of kidney disease. Unfortunately, interstitial fibrosis is poorly captured by the current clinical biomarkers of kidney function (eGFR and albuminuria). Thus, interstitial fibrosis is common, holds substantial prognostic importance, and yet clinicians are blind to its presence or severity except in rare instances when kidney biopsies are performed. Concurrently, new drugs are being tested to limit kidney interstitial fibrosis, but there are no non-invasive methods to assess changes in fibrosis over time. Interstitial fibrosis is currently estimated from histopathological examination of a kidney biopsy, which are rarely done. A non-invasive test to estimate interstitial fibrosis is not currently available. Our exciting preliminary data demonstrated that use of routine ultrasonography (USG) of the kidney, interpreted by deep learning/artificial intelligence can non-invasively assess the presence and severity of interstitial fibrosis. The overarching goal of this study is to further develop, and internally and externally validate a deep learning-based algorithm to estimate interstitial fibrosis from USG images of the kidney relative to the kidney biopsy gold standard. We hypothesize that, embedded within a kidney USG image are interstitial fibrosis corelates that can be extracted by deep learning and quantitatively analyzed to estimate interstitial fibrosis with high precision, and will improve prediction of longitudinal decline in kidney function. If so, given the widespread availability of kidney USG world-wide, this non-invasive estimate of interstitial fibrosis would have immediate clinical implications with improved prognostication, and ability to serially monitor interstitial fibrosis in response to therapy. The proposed program of research will address three specific aims: Aim 1. To further develop and internally validate a deep learning- based system for interstitial fibrosis quantification from kidney USG image. In Aim 2, we will externally validate the performance of the deep learning model using an independent cohort of USG images and kidney biopsies, and evaluate performance across strata of age, gender, and body size. Finally, in Aim 3, we will determine if the USG deep learning-based interstitial fibrosis score is associated with kidney disease progression with similar strengths relative to kidney biopsy assessment of interstitial fibrosis. Upon completion of this program of research, we envision development of an app. / plug-in for ultrasound reading modules that would facilitate widespread dissemination of the deep-learning tool, such that USG-based fibrosis scoring is widely available to treating clinicians.