Development and Dissemination of KiNet: A Novel Imaging Informatics Tool for Gastrointestinal and Pancreatic Neuroendocrine Tumors

项目摘要 神经内分泌肿瘤（NETs）是影响大多数器官系统的一种异质性癌症。NET必须 正确分级，以确保适当的治疗和病人管理。增殖指数， Ki 67核染色，是胃肠道（GI）和胰腺NET分级所需的， 世界卫生组织（卫生组织）。从病理学图像测量Ki 67标记指数（Ki 67 LI）需要 免疫阳性肿瘤、免疫阴性肿瘤和非肿瘤细胞的准确定量。这个过程是一个 基础、转化和临床研究以及常规临床实践中的基本程序。但目前的 Ki 67图像分析工具具有许多缺点：1）Ki 67 LI评估仍然主要通过以下方式实现： 手动或半自动化方法，导致劳动力成本增加，工作流程笨拙，吞吐量低 图像分析和显著潜在观察者间和观察者内变异性; 2）计算机辅助Ki 67计数， 由于多阶段图像处理设计容易出错，其中每一阶段本身就是一项非常具有挑战性的任务; 3)当前的算法设计没有考虑Ki 67图像的特征， 在Ki 67染色图像中分类不同类型细胞的技术困难。在这项研究中，我们 寻求开发和传播一种新的基于深度学习的成像信息学系统KiNeT， 用于更好地自动化GI和胰腺NET中的Ki 67 LI测量。KiNet将利用尖端的 机器学习算法，深度全卷积网络（FCN），以开发端到端，像素到像素的 单阶段Ki 67 LI评估模型。为此，我们将首先制定Ki 67计数作为细胞鉴定 问题，并在一个新的FCN网络中使用类感知结构化回归建模来解决它。这个网络 将同时检测和分类免疫阳性肿瘤、免疫阴性肿瘤和非肿瘤细胞。下一个是 我们将通过另一项相关任务--提取感兴趣区域（ROI）--进一步增强细胞识别， 将通过考虑Ki 67图像特征区分肿瘤和非肿瘤区域。这些 两个任务将被统一到一个单一的神经网络中，并联合学习，以利于细胞识别和 区域分类KiNet将为准确的Ki 67 LI评估提供一种新的计算方法， 从而能够对疾病进行早期检测和靶向治疗。与手工计数和电流相比， Ki 67图像分析方法，它将显著提高客观性、一致性、可靠性、重现性 效能此外，所提出的单级Ki 67计数策略，这是完全不同的， 目前的多级Ki 67图像分析管道，将为Ki 67图像量化提供新的视角。

PROJECT SUMMARY Neuroendocrine tumors (NETs) are one heterogeneous type of cancer affecting most organ systems. NETs must be correctly graded to ensure proper treatment and patient management. The proliferation index, as measured by Ki67 nuclear staining, is required for gastrointestinal (GI) and pancreatic NET grading per the criteria of the World Health Organization (WHO). Measuring the Ki67 labeling index (Ki67 LI) from pathology images requires accurate quantification of immunopositive tumor, immunonegative tumor and non-tumor cells. This process is an essential procedure in basic, translational and clinical research and in routine clinical practice. However, current Ki67 image analysis tools have a number of drawbacks: 1) Ki67 LI assessment is still mainly achieved with manual or semi-automated methods, leading to increased labor costs, awkward workflows, low-throughput image analysis and significant potential inter- and intra-observer variability; 2) computer-aided Ki67 counting is error-prone due to the multi-stage image processing design, where each stage itself is a very challenging task; 3) current algorithm design does not take into consideration the characteristics of Ki67 images such that it has technical difficulty in classifying different types of cells in Ki67 stained images. In this proposed research, we seek to develop and disseminate a novel deep learning-based imaging informatics system, KiNeT, specifically for better automated Ki67 LI measurement in GI and pancreatic NETs. KiNet will take advantage of cutting-edge machine learning algorithms, deep fully convolutional networks (FCNs), to develop an end-to-end, pixel-to-pixel model for single-stage Ki67 LI assessment. To this end, we will first formulate Ki67 counting as a cell identification problem and solve it using class-aware structured regression modeling within a novel FCN network. This network will simultaneously detect and classify immunopositive tumor, immunonegative tumor and non-tumor cells. Next, we will further enhance cell identification with another related task, extraction of regions of interest (ROIs), which will differentiate tumor from non-tumor regions by taking Ki67 image characteristics into consideration. These two tasks will be unified into one single neural network and jointly learned to benefit both cell identification and region classification. KiNet will provide a novel computational method for accurate Ki67 LI assessment, thereby enabling early detection and targeted treatments of the diseases. Compared to manual counting and current Ki67 image analysis methods, it will significantly improve the objectivity, consistency, reliability, reproducibility and efficiency. Additionally, the proposed single-stage Ki67 counting strategy, which is completely different from current multi-stage Ki67 image analysis pipelines, will provide a new perspective for Ki67 image quantification.