Computational Imaging of Renal Structures for Diagnosing DiabeticNephropathy

项目摘要 按照目前的速度，到2050年，三分之一的美国成年人将患有糖尿病。糖尿病继发的疾病是糖尿病。 肾病(DN)，导致美国22.5万名终末期肾病(ESRD)患者(占所有ESRD的50% 病例)，占每个患者每年医疗保险费用的1.9万美元。微量尿白蛋白的测定 (微量白蛋白尿)是糖尿病肾病最常见的非侵入性临床生物标志物。为了最终定义 糖尿病肾病的严重性，病理学家对肾脏中的肾小球结构损害进行定性的人工评估 活组织检查。然而，肾病活检中的肾小球结构通常与侵袭性较小的临床并不相关。 生物测定(例如，估计的肾小球滤过率、尿蛋白、血肌酐和血糖水平)。这 传统的诊断方法是近似的，受用户偏见的影响，耗时长，诊断率低 疾病早期阶段的准确性；此外，手动识别的特征可能并不总是准确预测 疾病的发展。计算图像分析提供了将临床生物特征投影到 肾小球组织学结构。此方法可更精确地识别导致 生理变化，从而减少了诊断所需的临床资源和时间，以及 为临床医生提供更大的反馈，以改善早期干预。我们已经开发了计算工具 目的：量化人类糖尿病肾病活检组织中的肾脏结构。我们的工具可量化组织学肾脏的肾小球特征 组织图像比人工方法更高效。我们还得出了一个量化进展风险分数 (PR)描述仅从单个活检点估计的糖尿病肾病进展风险。在这里，我们将严谨地分析 利用人糖尿病肾病肾脏的组织学图像预测疾病进展的这些方法的性能 活组织检查。我们将通过计算量化不同形态的、指示糖尿病肾病的肾小球内特征。我们 将分析地将计算得出的肾小球特征与临床生物识别相结合，以开发 患者特定的PR，以识别有肾功能衰竭风险的患者。由于人类肾脏的糖尿病肾病数据稀少，我们还将 使用鼠类数据，可以以受控的方式大量生成，以初始训练 计算模型。然后，我们将通过使用人类对参数进行微调来改进模型以供临床使用 数据。创新之处在于将传统的临床检测方法与传统的诊断方法进行了新的结合 方法，在提高精度的计算模式下。这种集成将导致计算 疾病预测最早可测量的肾功能不全的生物标志物。我们将研究可预测的 这些标记物从早期时间点预测未来临床终点的能力。这些方法支持 发展可量化的预测和预测信息，这些信息在疾病过程中是动态的， 易于辨别，对于模拟疾病进展或对治疗的反应具有很高的信息量。这 研究将1)实现更早的临床预测，从而延长糖尿病肾病进展的干预窗口；以及2) 作为未来研究的试点平台，通过计算得出预测其他疾病的肾脏生物标志物。

Project Summary At the current rate, one in three U.S. adults will be diabetic by 2050. A disease secondary to diabetes is diabetic nephropathy (DN), which causes end-stage renal disease (ESRD) for >225K U.S. patients (50% of all ESRD cases), accounting for >$19K in yearly Medicare costs for each patient. Measurement of minute urinary albumin (microalbuminuria) is the most common non-invasive clinical biomarker of DN. In order to conclusively define DN severity, pathologists conduct qualitative manual estimation of glomerular structural damage in renal biopsies. However, renal glomerular structure in DN biopsies does not often correlate with less invasive clinical biometrics (e.g., estimated glomerular filtration rate, urine protein, serum creatinine and glucose levels). This traditional diagnostic method is approximate, subjected to user bias, time-consuming, and has low diagnostic precision in early disease stages; further, manual hand identified features may not always accurately predict disease progression. Computational image analysis offers the opportunity to project clinical biometrics onto glomerular histological structures. This method provides finer precision in identifying structural changes that lead to physiological changes, which in turn reduces the required clinical resources and time for diagnosis, and provides clinicians with greater feedback to improve early intervention. We have developed computational tools to quantify renal structures in human DN biopsies. Our tools quantify glomerular features in histological renal tissue images more efficiently than manual methods. We have also derived a quantitative progression risk score (PRS) describing DN progression risk estimated off only a single biopsy point. Here, we will rigorously analyze the performance of these methods to predict disease progression using histological images of human DN renal biopsies. We will computationally quantify morphologically diverse DN-indicative intra-glomerular features. We will analytically integrate computationally derived glomerular features with clinical biometrics in order to develop patient-specific PRS to identify patients at risk of renal failure. Since human renal DN data is sparse, we will also use murine data, which can be generated in large amounts in a controlled fashion, to initially train the computational models. We will then refine the model for clinical use by fine-tuning the parameters using human data. The innovation is in the novel integration of traditional clinical detection methods with traditional diagnostic methods, under a computational schema for enhanced precision. This integration will lead to computational disease predicting biomarkers of the earliest measurable renal DN dysfunction. We will study the predictive power of these markers to foretell future clinical endpoints from earlier time points. These methods support the development of quantifiable prognostic and predictive information, which is dynamic over the disease course, easily discriminated, and is highly informative for modeling disease progression or response to therapy. This study will 1) enable earlier clinical predictions, thus extending windows for interventions of evolving DN; and 2) work as a pilot platform for future studies to computationally derive renal biomarkers predictive of other diseases.