A Model-based Approach to Comparing Breast Images

乳腺X线摄影是筛查乳腺癌的首选方法。乳房X线照片是乳房的X射线图像，在乳房被紧密压缩在近似平行的板之间时拍摄。通常使用两种不同的视图：头尾位（CC /垂直向下看乳房）和内外斜位（MLO /从对侧肩部看乳房）。在少数情况下，可以在一次乳房X光检查中检测到潜在的癌症迹象，但更常见的是，放射科医生通过比较乳房X光检查来检测异常。在双侧比较中，放射科医生比较左右乳房的乳房X线照片，并寻找图像之间的不对称和结构扭曲，尽管这些术语定义不清（因此不容易构建到算法中）。在时间比较中，放射科医生将当前乳房X线照片与前一张进行比较，并寻找显著变化的迹象。在多视图比较中，放射科医生比较同一乳房的CC和MLO视图，并寻找异常的一致证据。通过经验，放射科医生学会了进行这些比较，但这项任务本质上是困难的，因为生物变异，软组织变形和投影成像的影响是混淆的，因此不可能在乳房X线照片之间建立逐点对应。我们的目标是开发计算方法，以协助放射科医生在这种比较。我们计划采取一种全面的基于模型的方法来进行有意义的比较。这将借鉴我们以前的图像配准的经验，以及模拟解剖变异性的统计数据，组织变形的生物力学和图像形成的物理学。为了了解这些不同来源的变异性在2D乳腺X线照片中的影响，并开发适当的模型，我们将利用大量现有的3D MR乳腺图像和相应的乳腺X线照片。乳腺成像的应用本身就有价值，有可能为更有效的计算机辅助检测系统做出重大贡献，但我们提出开发的方法是通用的，能够广泛应用于医学图像分析中的其他应用，其中器官在经历机械变形的同时成像/例如，心脏，肝脏，肺，胃和结肠。

X-ray mammography is the method of choice in screening for breast cancer. Mammograms are x-ray images of the breast, taken with the breast tightly compressed between approximately parallel plates. Two different views, are commonly used: cranio-caudal (CC / looking vertically down on the breast) and medio-lateral oblique (MLO / looking at the breast from the opposite shoulder). In a few cases, potential signs of cancer can be detected in a single mammogram, but more usually, radiologists detect abnormalities by comparing mammograms. In bilateral comparison, the radiologist compares the mammograms of the left and right breasts and looks for asymmetries and architectural distortions between the images, though these terms are ill-defined (and thus not easily built into an algorithm). In temporal comparison, the radiologist compares the current mammogram with the previous one and looks for signs of significant change. In multi-view comparison, the radiologist compares the CC and MLO views of the same breast and looks for consistent evidence of an abnormality. Through experience, radiologists learn to make these comparisons, but the task is intrinsically difficult, because the effects of biological variation, soft tissue deformation, and projection imaging are confounded, so that establishing point-by-point correspondence between mammograms is impossible. Our aim is to develop computational methods for assisting the radiologist in this comparison. We plan to take a comprehensive model-based approach to making meaningful comparisons. This will draw on our previous experience of image registration together with modelling the statistics of anatomical variability, the biomechanics of tissue deformation, and the physics of image formation. To understand the effects of these different sources of variability in 2D mammograms, and develop appropriate models, we will make use of a large existing collection of 3D MR breast images, with corresponding mammograms. The application to breast imaging is of value in its own right, with the potential to make a significant contribution to more effective systems for computer-aided detection, but the methods we propose to develop are generic and capable of broad application to other applications in medical image analysis, where organs are imaged whilst undergoing mechanical deformation / for example, the heart, liver, lung, stomach and colon.