TRD 2 - Deep Learning

国家信息技术中心先进技术深度学习技术研发的目标 图像引导治疗(AT-NCIGT)将研究深度学习(DL)方面的革命性进展 脑癌、前列腺癌和肺癌的图像引导治疗(IGT)的背景，并开发工具 可供更广泛的IGT研究社区使用。我们研究的总主题是 与创建训练数据有关的困难，这是应用的一个重大障碍 医学影像的数字图书馆。而DL在基于图像的分类或分割方面已经取得了许多成功 任务，所使用的方法是完全受监督的，即非常大量的准确注释的培训 需要数据才能实现最佳性能。目前，专家标注在以下方面既昂贵又费力 医学图像的情况，因为3D结构的准确分割需要人工或 每幅图像可半自动标记数千个体素。同时，大型未标记或 弱标记的数据集正在变得可用。例如，某大医院的PACS系统 可能包含数千万次扫描，但确定准确的疾病标签是困难的。确实有 目前有两种很有前途的动态链接法可以用来解决这个问题，它们是弱监督的 学习(在某些标签缺失或不完美的情况下)和迁移学习(利用 在某些方面相似的标记数据集)。目前的情况因缺乏资金而进一步恶化 机器可读的元数据，以及用于支持成像的管理的方法和工具(例如，磁的 磁共振成像(MRI))和临床数据，以及注释和分析结果 数据模型。后者导致数据碎片化，以及非标准和异构元数据。 我们通过1)开发新的信息理论技术来解决这些问题 弱监督深度学习，2)开发癌症深度学习的新训练策略 会阴孔内MRI引导的前列腺活检的特征，以及3)开发基础设施 用于为深度学习整理成像数据。此TRD的结果将是DL算法，结果是 模型以及用于批注和组织机器可读元数据的工具，这些模型和工具旨在 IGT癌症研究适用于前列腺癌、肺癌和脑部。

The goals of the Deep Learning TRD of the Advanced Technologies for the National Center for Image-Guided Therapy (AT-NCIGT) are to investigate revolutionary advances in deep learning (DL) in the context of image-guided therapy (IGT) of brain, prostate, and lung cancer, and to develop tools that can be used by the broader IGT research community. The general theme of our research addresses difficulties associated with creation of training data, which is a significant impediment to the application of DL to medical images. While DL has had many successes in image-based classification or segmentation tasks, the methods used are fully supervised, i.e., very large amounts of accurately annotated training data are needed to achieve best performance. Currently, expert annotation is expensive and laborious in the case of medical images because accurate segmentation of 3D structures requires manual or semi-automatic labeling of thousands of voxels per image. Concurrently, large unlabeled or weakly-labeled data sets are becoming available. For example, the PACS system of a large hospital might contain tens of millions of scans, but determining accurate disease labels is difficult. There are currently two promising DL approaches that can be used to address this problem, weakly-supervised learning (where some labels are absent or otherwise imperfect) and transfer learning (which leverages labeled data sets that are in some ways similar). The current situation is further exacerbated by a lack of machine readable metadata, and of methods and tools to support curation of the imaging (e.g., Magnetic Resonance Imaging (MRI)) and clinical data, alongside annotations and analysis results within a single data model. The latter leads to fragmentation of data, and non-standard and heterogeneous metadata. We address these problems by 1) Developing new information theoretic technology for weakly-supervised deep learning, 2) developing novel training strategies for deep learning for cancer characterization for transperineal in-bore MRI-guided prostate biopsy, and 3) developing an infrastructure for curating imaging data for deep learning. The results of this TRD will be DL algorithms, the resulting models, and tools for annotation and organizing machine-readable metadata that are designed to enable IGT cancer research for the prostate, the lung, and the brain applications.