Case-level detection of mammographic masses

Resumen

This thesis is focused on the automatic detection of masses in FFDM images by using case-level information which includes bilateral, temporal and/or ipsilateral information. As a first step, FFDM images are preprocessed to improve image quality before working on the proper detection framework. A novel enhancement method is applied to compensate the thickness reduction in peripheral edges of the breast in FFDM. Following, B-Splines image registration with Affine initialisation is used to obtain bilateral and temporal information that is incorporated in the detection stage of the whole process. This registration approach is considered the optimal one that provides useful and usable case-level information among several investigated registration methods. Finally, CC/MLO correspondence approach based on using curved epipolar lines is used in the FP stage. Ipsilateral information allows to distinguish between real (when CC/MLO lesion correspondence exists) and non-real (when there is no CC/MLO lesion correspondence) masses. Furthermore, in order to add breast density information to the detection process, different methods for breast density assessment are analysed. Both, qualitative and quantitative methods are proposed and evaluated. Initial results show a better performance of the multi-image CAD approach relative to the single-image CAD approach. Sensitivity increases and the number of FPs is reduced. 29