Deep learning and topological approaches to identify kidney tissue features associated with adverse outcomes after nephrectomy

摘要 肾活检组织的病理学评估仍然是肾移植患者不良结局的最佳预测因子。 肾脏疾病这些特征在很大程度上与疾病病因无关，并且在非疾病中没有很好地反映出来。 侵入性检查（如血清肌酐和白蛋白尿）。定量评估这些参数是时间 消耗并且可能由于肾组织内的病理特征的异质性而缺陷。我们建议 评估和优化计算机图像分析方法，以支持大块的病理分析 我们收集了220例接受肾切除术的患者的无癌肾组织。 对这些样本中肾小球表型的计算机辅助分析显示， 无明显病理改变的肾小球先于确定的病理改变。我们假设评估 无癌肾组织的转移将告知剩余肾脏的亚临床损伤， 相关的病理和临床参数。我们建议评估肾小球、动脉和肾小管， 确定所述肾组织内所评估的特征的空间相互关系。 对比针吸活检获得的肾组织块明显更大的肾组织块的检查允许 包括20倍肾小球（肾切除术样品：AVRG. 256个肾小球/样本;针吸活检：avrg。 13/样品），根据标准病理学标准，绝大多数被认为是"表现正常"。在 此外，这些样品包括大量的血管（肾切除样品：AVRG，18 动脉/样本;针吸活组织检查：平均值。1/样品），从而能够更稳健地评价血管系统。我们 建议应用并优化我们的检测和分割方法来检测肾小球、动脉和 管状段来训练卷积神经网络，并使用拓扑图像分析来自动化 识别视觉和亚视觉特征。此外，我们还将评估 单个特征（肾小球、动脉和肾小管段以及相同类别的特征，即全局 硬化的肾小球，玻璃样变性的动脉，萎缩的小管）。为了确定重现性， 在我们的方法中，我们将评估来自相同样本的单独部分的第二个组织切片。具体地说， 我们提出了一种使用深度学习的肾脏特征的算法检测和表征， 用于发现肾组织图像中新亚视觉特征并确定空间 这些特征的相关性。 如果成功，我们将在未来的独立研究中验证我们的分析方法。为此，我们 已经前瞻性地收集了来自同意接受治疗的患者的肾组织和纵向临床数据， 肾切除术，允许将特定特征与临床相关结果相关联。

ABSTRACT Pathologic assessment of kidney biopsy tissue remains the best predictor of adverse outcomes in patients with kidney diseases. These features are largely independent of disease etiology and are not well reflected in non- invasive tests (e.g. serum creatinine and albuminuria). Quantitative assessment of these parameters is time consuming and maybe flawed by heterogeneity of pathologic features within kidney tissue. We propose to evaluate and optimize computational image analysis approaches to support pathologic analysis of large pieces of cancer-free kidney tissue from patients who underwent nephrectomy which we have collected (n > 220). Computer-assisted analysis of glomerular phenotypes in these samples show that morphometric features in glomeruli without obvious pathology precede established pathologic changes. We hypothesize that evaluation of cancer-free kidney tissue will inform about subclinical damage in the remaining kidney which is associated with relevant pathologic and clinical parameters. We propose to assess glomeruli, arteries and tubuli, and determine the spatial inter-relationship of the assessed features within the kidney tissue. The examination of significantly larger pieces of kidney tissue than those obtained by needle biopsy allows to include 20 times more glomeruli (nephrectomy samples: avrg. 256 glomeruli/sample; needle biopsy: avrg. 13/sample) with the vast majority considered “normal appearing” as per standard pathologic criteria. In addition, these samples include a significant larger number of blood vessels (nephrectomy samples: avrg. 18 arteries/sample; needle biopsy: avrg. 1/sample) allowing a more robust evaluation of the vasculature. We propose to apply and optimize our detection and segmentation approach to detect glomeruli, arteries and tubular segments to train convolutional neural networks and use topological image analysis to automate the identification of visual and sub-visual features. In addition, we will assess the spatial relationship between individual features (glomeruli, arteries and tubular segments and features of the same category, i.e. globally sclerosed glomeruli, arteries with hyalinosis, atrophied tubuli) within the section. To determine reproducibility of our approach, we will assess a second tissue section from a separate part of the same samples. Specifically, we propose an algorithmic detection and characterization of kidney features using deep learning, a topological image analysis for discovery of novel sub-visual features in kidney tissue images and to determine spatial relatedness of these features. If successful, we will validate our analytical approach in future independent studies. For this purpose, we are already prospectively collecting kidney tissue and longitudinal clinical data from consented patients undergoing nephrectomies, allowing association of specific features with clinical relevant outcomes.