Characterising Features of Breast Tissue using Radiodensity Measured in X-Ray Mammograms and Tomosynthesis for Cancer Risk Assessment and Diagnosis

自20世纪60年代末以来，在称为乳房密度的研究领域中，一直在研究乳房内组织结构的放射学特征与乳房包含或随后发展为癌症的可能性之间的相关性。在癌症风险预测方面，已发表的许多研究表明，在乳房X光摄影图像中观察到的某些放射学组织特征与受试者最终发展为乳腺癌之间存在统计上的显著相关性。最近人们对这类技术的兴趣激增，因为人们越来越希望更具体地针对一般筛查计划的巨大任务。这不仅是为了减轻经济负担，也是为了确保高危人群得到最佳护理，特别是通过使用其他(更昂贵的)成像手段，如增强磁共振成像，以补充X光乳房X光检查。遗憾的是，许多现有的乳房密度评估技术都有许多缺点：通常它们依赖于读者对不确定特征集的视觉评估，因此严重受制于读者的主观性；并且几乎普遍地忽略了用于采集的X射线的特征对图像内组织外观的影响。在我现有的图像形成过程建模工作的坚实基础上，特别是包括散射辐射，以及我们首创的辐射密度的标准衰减率测量，我们的目标是建立一种完全自动化的评估技术，该技术考虑到图像采集参数的影响，因此仅依赖于存在的组织。使用最先进的计算机学习方法、我们的临床合作者两年来获得的所有数字乳房X光照片的数据库，以及将在项目期间迭代开发的特征描述向量，将进行彻底和详细的分析，以尽可能具体地确定乳房内可能发生癌症的潜在组织特征。这项工作还将使用同样的方法，扩展到计算机检测恶性肿瘤，包括微钙化和病变，希望提高现有对比特征识别算法的灵敏度和特异度，并将进一步将开发的技术应用于三维乳房X光摄影，特别是断层合成。三维乳房X光摄影提供了消除特征噪声的巨大好处：占据乳房不同垂直位置的几个结构在投影图像中彼此重叠，从而看起来像一个单一实体的复杂情况。这已经被发现通过掩盖被放射致密组织包围的恶性肿瘤来降低乳房X光检查的敏感性。在这项工作中还开发了成像模型的应用，特别是在从源投影图像重建三维体数据时消除使用物理栅格反散射装置的有害剂量影响的散射的软件校正。

Since the late 1960s the correlation between radiological features of the tissue structures within the breast and the likelihood of the breast containing, or subsequently developing cancer has been studied in the field of research termed breast density. In terms of cancer risk prediction, many studies have been published showing statistically significant correlations between certain radiographic tissue characteristics observed within a mammographic image, and the subject ultimately developing breast cancer. Recent interest in such techniques has surged as the desire grows to more specifically target the huge undertaking of a general screening programme. This is not only to reduce economic burden, but to ensure those at high risk get optimal care, particularly through the use of other (more expensive) imaging modalities, such as contrast enhanced MRI, to supplement the x-ray mammogram. Unfortunately many of the existing breast density assessment techniques have a number of short comings: generally they rely on reader visual assessment of an uncertain feature set, and so badly suffer from reader subjectivity; and almost universally the effect of the characteristics of the x-rays used for acquisition on tissue appearance within the image are ignored. Building upon the firm foundation of my existing work in modelling the image formation process, in particular the inclusion of scattered radiation, and the standard attenuation rate measure of radiodensity we pioneered, we aim to build a fully automated assessment technique, which takes account of the effects of the image acquisition parameters and therefore solely depends on the tissue present. Using state of the art computer learning methods, a database of all the digital mammograms acquired over two years by our clinical collaborator, and a feature description vector that will be iteratively developed during the project, a thorough and detailed analysis will be performed to establish as specifically as possible the underlying tissue features within a breast likely to develop cancer. The work will also extend, using the same methodology, to the computer detection of malignancies, in the form of both microcalcifications and lesions, in the hope of improving the sensitivity and specificity of the existing contrast feature recognition algorithms.Further the application of the developed techniques to three-dimensional mammography, in particular tomosynthesis will be undertaken. Three-dimensional mammography gives the enormous benefit of removing feature noise : the complication of several structures occupying different vertical positions in the breast becoming superimposed upon each other within the projection image and thereby appearing as a single entity. This has been found to decrease mammographic sensitivity by masking malignancies surrounded by radiographically dense tissue. The application of the model of image formation, and in particular the software correction of scatter which removes the harmful dose implications of the use of physical grid anti-scatter devices, upon the reconstruction of the three-dimensional volume data from the source projection images is also being developed in this work.