INVESTIGATIONS OF MULTI-VIEW CAD FOR MAMMOGRAPHY

在过去的15年里，在计算机辅助检测(CAD)乳房X光检查异常方面的发展取得了长足的进步。然而，由于当前CAD算法的性能限制，CAD是否提供净收益的问题仍然没有解决。近年来，尽管许多组织做出了相当大的努力，但CAD性能的改善速度已经下降到性能统计似乎正在接近渐近线的程度，这远远低于乳房X光师的性能。最有可能的原因是，目前所有的CAD实现基本上都建立在传统的信号处理和模式识别方法之上，并通过检测单个图像中的特征来获得它们的性能。然后通过某种推理机制对这些特征进行分类。可以想象(很可能)，单个图像中的大多数相关物理特征已经在某种程度上被识别和利用。这一建议的假设是，当前CAD的性能限制，如CAD和乳房X光摄影者之间的性能差异所表明的，在很大程度上是由于这些算法未能利用只能通过多个图像的协同分析得出的数据。因此，这项建议的目的是扩展现有的CAD方法，使之能够从同侧视图中提取与乳房空间结构相关的信息。我们的初步结果表明，尽管压缩引起了失真，但仍有一些特征可以从成对的图像中自动得出，并且已经被证明提供了从对单个图像的独立分析中无法获得的信息。这些基于多图像的特征在一定程度上独立于乳房X光摄影过程中乳房压缩造成的组织变形。我们将调查和完善这些和其他可以识别的特征。本研究的目的是充分利用这些特征，并优化它们对基于多图像的CAD算法的贡献。

Over the past 15 years, considerable progress has been made in the development of computer-aided detection (CAD) of abnormalities in mammograms. Nevertheless, because of performance limitations of current CAD algorithms, the question of whether CAD provides a net benefit remains unresolved. In recent years, despite considerable effort by many groups, the rate of improvement in CAD performance has declined to the point that performance statistics seem to be approaching an asymptote, which is well below the performance of mammographers. The most likely reason for this is that essentially all current CAD implementations are founded on traditional methods of signal processing and pattern recognition and derive their performance by detecting features in a single image. These features are then classified by some inference mechanism. It is conceivable (probable) that most of the relevant physical features in single images have been identified and exploited to some extent. The hypothesis of this proposal is that performance limitations of current CAD, as indicated by the difference in performance between CAD and mammographers, result to a large extent from the failure of these algorithms to utilize data that can only be derived by a synergistic analysis of multiple images. Thus, it is the intent of this proposal to extend current CAD methodology to enable the extraction of information related to the spatial structure of a breast from ipsilateral views. Our preliminary results have established that despite compression-induced distortion, there are features that can be derived automatically from pairs of images and have been shown to provide information not obtainable from the independent analysis of single images. These multi-image-based features are partially independent of tissue distortion from breast compression during mammography. We will investigate and refine these and other features that can be identified. The purpose of this investigation is to fully exploit these kinds of features and optimize their contributions to a multi-image-based CAD algorithm.