Multimodality CAD system with image references for breast mass characterization

描述(由申请人提供)：该项目的长期目标是开发一种有效的计算机辅助诊断(CAD)系统，以帮助放射科医生在乳房成像中做出诊断决定。在这个拟议的项目中，我们将专注于使用乳房X光照片和超声图像来描述肿块的特征。我们提出了一种新的基于分类器的CAD方法，该方法可以同时估计肿块的恶性可能性，并从大量已知诊断的病例库中检索类似的病例，供放射科医生参考。因此，新的CAD系统结合了基于评级和基于图像检索的CAD系统的优点。它将帮助放射科医生不仅通过恶性评估，而且通过加强他们的基于相似的决策过程。我们还将设计一个相关反馈图像检索系统，使放射科医生能够交互和高效地从大型数据集中检索类似的病例，作为帮助开发自动化CAD系统的工具。我们假设，参考图像将提高经验较少的读者对大众的定性准确性，本研究开发的计算机分类和图像检索系统将显著提高放射科医生的准确性。为了验证这些假设，我们将执行以下具体任务：(1)收集包含肿块的超声图像和乳房X光照片的数据库；(2)提取用于描述肿块特征的特征；(3)开发决策树和k近邻分类器，比较使用和不使用Booost的决策树训练，并基于开发的分类器研究检索相似病例的方法；(4)开发相关反馈图像检索方法；(5)比较经验较少的放射科医生在没有和有经验的放射科医生检索的参考图像的帮助下的表现；以及(6)通过受试者工作特性(ROC)研究比较放射科医生在没有和使用全自动分类和图像检索CAD系统的情况下的表现。如果开发成功，CAD系统不仅可以减少良性活检，还可以减少经验丰富和经验较少的放射科医生之间在解释方面的差异。该项目与公共卫生的相关性在于，70%-85%的乳腺活检是针对良性病变进行的。在不降低乳腺癌检测灵敏度的情况下，这一数字的任何减少都将降低医疗保健成本，并通过减少焦虑和发病率来促进患者的福祉。

DESCRIPTION (provided by applicant): The long term goal of the project is to develop an effective computer-aided diagnosis (CAD) system to assist radiologists in making diagnostic decisions in breast imaging. In this proposed project, we will concentrate on the characterization of masses using mammograms and ultrasound images. We propose a new approach to CAD based on a classifier that can simultaneously estimate the likelihood of malignancy for the mass and retrieve similar cases from a large library of cases with known diagnosis for the radiologist's references. The new CAD system thus combines the advantages of a rating-based and an image-retrieval- based CAD system. It will aid radiologists not only by the malignancy estimate but also by enhancing their similarity-based decision making process. We will also design a relevance feedback image retrieval system that allows the radiologist to interactively and efficiently retrieve similar cases from a large data set as a tool to help develop the automated CAD system. We hypothesize that the reference images will increase the characterization accuracy of less experienced readers for masses, and that the computerized classification and image retrieval system to be developed in this study will significantly improve radiologists' accuracy. To test these hypotheses, we will perform the following specific tasks: (1) collect a database of sonograms and mammograms containing masses; (2) extract features for mass characterization; (3) develop decision tree and k-nearest neighbor classifiers, compare decision tree training with and without boosting, and investigate methods for the retrieval of similar cases based on the developed classifiers; (4) develop a relevance feedback image retrieval method; (5) compare the performances of less experienced radiologists without and with aid by reference images retrieved by experienced radiologists; and (6) compare radiologists' performances without and with the fully-automated classification and image-retrieval CAD system by a receiver operating characteristic (ROC) study. If successfully developed, the CAD system may not only reduce benign biopsies, but also reduce the variation in interpretation between experienced and less experienced radiologists. The relevance of this project to public health is that 70-85% of breast biopsies are performed for benign lesions. Any reduction in this number without a decrease in breast cancer detection sensitivity will decrease health care costs, as well as contribute to the well-being of the patient by reducing anxiety and morbidity.