Curvilinear Structure Based Mammographic Registration

Mammographic registration is a challenging problem due in part to the intrinsic complexity of mammographic images, and partly because of the substantial differences that exist between two mammograms that are to be matched. In this paper, we propose a registration algorithm for mammograms which incorporates junctions of Curvilinear structures (CLS) as internal landmarks. CLS depict connective tissue, blood vessels, and milk ducts. These are detected by an algorithm based on the monogenic signal and afforced by a CLS physical model. The junctions are extracted using a local energy (LE)-based method, which utilises the orientation information provided by the monogenic signal. Results using such junctions as internal landmarks in registration are presented and compared with conventional approaches using boundary landmarks, in order to highlight the potential of anatomical based feature extraction in medical image analysis. We demonstrate how computer vision techniques such as phase congruency (PC), local energy (LE) and multi-resolution can be applied in linear (1-D) and junction (2-D) detection as well as their application to medical image registration problems.