Interpretable Deep Learning Algorithms for Pathology Image Analysis

用于病理图像分析的可解释深度学习算法 摘要 染色组织的显微镜检查是生物医学研究的基本组成部分，对于 了解疾病的生物学过程，从而改善诊断、预后和治疗 响应预测。从癌症诊断到心脏排斥反应和法医主观解释 组织病理学切片构成临床决策和研究结果的基础。然而，它已经 已经表明，这种对病理切片的主观解释受到大量观察者和 观察者内部的可变性。计算机视觉和深度学习的最新进展使客观和 图像的自动分析。这些方法已成功应用于组织学图像，这些图像具有 展示了开发客观图像解释范例的潜力。然而，重要的是 在可以使用这种组织学图像的客观分析之前，算法方面的挑战仍然有待解决 由临床医生和研究人员提供。利用在开发和大量使用研究软件方面的丰富经验 基于深度学习，PI将开创新的算法方法来应对这些挑战，包括 但不限于：(1)使用千兆像素大小显微镜训练数据高效且可解释的深度学习模型 使用弱监督标签的图像分类和分割(2)数据的根本重新设计 用于整合来自显微镜图像和分子图谱的信息的融合范例(来自多组学 数据)用于改进诊断和预后判定(3)开发可视化和解释 用于研究人员和临床工作流程的软件，以提高临床和研究的有效性和重复性。 该系统将以模块化、用户友好的方式设计，并将是开放源码的，可通过 GitHub作为通用的即插即用模块，可适用于各种临床和研究应用。 我们还将开发一个网络资源，其中包含各种器官、疾病状态和亚型的预先训练的模型 这些将伴随着详细的手册，以便研究人员可以将深度学习应用于其特定的 研究问题。总体而言，该实验室的研究将从病理图像中产生影响很大的发现 分析，它的软件将使许多其他NIH资助的实验室能够做同样的事情，跨越各种 生物医学学科。

Interpretable Deep Learning Algorithms for Pathology Image Analysis Abstract The microscopic examination of stained tissue is a fundamental component of biomedical research and for the understanding of biological processes of disease which leads to improved diagnosis, prognosis and therapeutic response prediction. Ranging from cancer diagnosis to heart rejection and forensics the subjective interpretation of histopathology sections forms the basis of clinical decision making and research outcomes. However, it has been shown that such subjective interpretation of pathology slides suffers from large interobserver and intraobserver variability. Recent advances in computer vision and deep learning has enabled the objective and automated analysis of images. These methods have been applied with success to histology images which have demonstrated potential for development of objective image interpretation paradigms. However, significant algorithmic challenges remain to be addressed before such objective analysis of histology images can be used by clinicians and researchers. Leveraging extensive experience in developing and decimating research software based on deep learning the PI will pioneer novel algorithmic approaches to address these challenges including but not limited to: (1) training data-efficient and interpretable deep learning models with gigapixel size microscopy images for classification and segmentation using weakly supervised labels (2) fundamental redesign of data fusion paradigms for integrating information from microscopy images and molecular profiles (from multi-omics data) for improved diagnostic and prognostic determinations (3) developing visualization and interpretation software for researchers and clinical workflows to improve clinical and research validation and reproducability. The system will be designed in a modular, user-friendly manner and will be open-source, available through GitHub as universal plug-and-play modules ready to be adapted to various clinical and research applications. We will also develop a web resource with pretrained models for various organs, disease states and subtypes these will be accompanied with detailed manuals so researchers can apply deep learning to their specific research problems. Overall, the laboratory’s research will yield high impact discoveries from pathology image analysis, and its software will enable many other NIH funded laboratories to do the same, across various biomedical disciplines.