III-CXT-Small: Collaborative Research: Structuring, Reasoning, and Querying in a Very Large Medical Image Database

图像数据在医学信息学中具有重要的实用价值，也是工业界和学术界研究人员非常感兴趣的课题。虽然数字图像数据库现在在临床和教育环境中很流行，传统的与这种集合交互和查询的方法可以提供某种程度的有用功能，但试图弥合语义鸿沟的系统例子很少。这项拨款建议的工作是一项多机构合作，将医学图像处理、机器学习和模式识别、知识表示和查询以及该领域专家的评估研究结合在一起，旨在推动这一方向的最新进展。由国家医学图书馆和国家癌症研究所收集的60,000张宫颈X光照片档案是实现这一目的的理想收藏。NLM宫颈造影档案形成了一个狭窄的图像域，具有有限和可预测的可变性。在这种情况下，领域知识的显式表示缓解了场景的低级感觉记录(原始图像数据)与图像中隐含的对象和过程(语义解释)之间的语义鸿沟。这项研究将遵循从原始图像数据到低级别图像特征、物体和组织类型的识别、关于疾病过程的基于知识的推理、最后是支持临床和NLM/NCI合作者诊断决策的工具和可视化的信息层次结构。研究团队将采用一种名为计算机辅助视觉交互识别(CAVIAR)的基本范式，该范式将领域专家视为等式的组成部分，并试图优化整个人机系统的性能。知识价值图像内容理解仍然被认为是一个令人烦恼的开放问题，与此同时，数据库的大小和复杂性正在迅速增长。预计这项工作将对与医学图像分析有关的领域产生积极影响，包括信息提取、组织、表示和查询，以及培训。通过对NLM/NCI宫颈造影术档案的关注，这项研究可能有助于推动宫颈造影术作为一种比宫颈涂片和阴道镜更具成本效益的程序在宫颈癌筛查中的作用。这一目标领域项目的结果还可以揭示差距，并有助于为更广泛领域的研究确定新的优先事项，例如多媒体内容的结构、理解、索引和检索。

Image data is of immense practical importance in medical informatics, and a subject of strong interest to researchers in industry and academia. While digital image databases are now prevalent in clinical and educational settings, and traditional means for interacting with and querying such collections can provide some level of useful functionality, there are few examples of systems that attempt to bridge the ?semantic gap.? The work proposed in this grant is a multi-institutional collaboration combining research in medical image processing, machine learning and pattern recognition, knowledge representation and querying, and evaluation by domain experts in the field, is intended to advance the state-of-the-art in this direction. The archive of 60,000 cervigram images assembled by the National Library of Medicine and National Cancer Institute is an ideal collection for this purpose. The NLM cervigram archive forms a narrow image domain that has a limited and predictable variability. In such cases, explicit representation of domain knowledge alleviates the semantic gap between the low-level sensory recordings of a scene (raw image data), and objects and processes implied from images (semantic interpretation). This research will follow an information hierarchy that proceeds from raw image data to low-level image features, recognition of objects and tissue types, knowledge-based reasoning about disease processes, and, finally, tools and visualizations to support diagnosis decisions by clinical and NLM/NCI collaborators. The research team will employ an underlying paradigm known as Computer-Assisted Visual Interactive Recognition, or CAVIAR, which considers the domain expert an integral part of the equation and attempts to optimize the performance of the complete human-machine system. Intellectual Merit Image content understanding is still considered a vexing open problem at the same time databases are growing rapidly in size and complexity. It is anticipated that this work will have a positive impact in areas relating to medical image analysis, including information extraction, organization, representation, and querying, as well as in training. Broader Impact Through the focus on the NLM/NCI cervigram archive, this research may help advance the role of cervicography as a more cost-effective procedure than pap smears and colposcopy in screening for cervical cancer. Results from this targeted-domain project could also illuminate gaps and help establish new priorities for research in broader domains such as multimedia content structuring, understanding, indexing, and retrieval.