Tracking Mammographic Structures Over Time

A method to correspond linear structures in mammographic images is presented. Our approach is based on automatically establishing correspondence between linear structures which appear in images using robust features such as orientation, width and curvature extracted from those structures. The resulting correspondence is used to track linear structures and regions in mammographic images taken at different times. Medical image analysis [1] has been an important research subject in recent years where computer vision techniques have been successfully applied to develop detection and diagnosis systems, enhancement and training tools. The analysis of mammographic images is one of those fields and as such a very challenging one due to the complexity of the images and the subtle nature of the abnormalities. Detection of abnormal structures or architectural distortions in mammographic images can be performed by analysing different images of the same patient. Various approaches have been adopted which bring images into alignment in order to detect differences which are likely to be due to an abnormality. A large number of those methods are based on automatically corresponding extracted landmarks from mammographic images. Those landmarks include breast boundary [15, 20, 9], pectoral muscle [9], salient regions extracted using wavelets [14], iso-intensity contours [11] or steerable filters [20] and crossing points of horizontal and vertical structures [22]. This work presents an approach to the correspondence in mammographic images based on anatomical features which appear as linear structures in the images. The correspondence is used here to track linear structures in mammograms of the same patient over several years. Tracking of linear structures could be used to assess and model the development of architectural changes and abnormal structures. By being able to track regions back in time the available information will help to improve early detection of subtle abnormalities which are initially missed by radiologists. The tracking of objects in image sequences is a well-developed area [21]. However, in general this involves rigid objects (like cars [8]) or objects with a predictable behaviour (like humans [10] or animals [18]). Another difference with the current application is the fact that normally tracking is established using sequences of tens to hundreds of images and not only a few.