Document Overview

The thesis is structured according to the mentioned objectives. Thus, the following chapter is an extensive survey of the mass segmentation methods found in the literature. A qualitative and quantitative analysis is performed, extracting the corres-

Figure 1.5: Scheme of our proposal.

ponding conclusions. As already explained, these supported the influence of breast tissue and the lesion shape and size on the performance of the algorithms.

Chapter  covers the classification of the breasts according to their internal tissue. It starts with a brief survey of the few existing methods dealing with this issue and subsequently, a new algorithm is proposed and exhaustively evaluated using two different databases and the opinion of three different radiologists.

Chapter  describes the developed algorithm for finding masses in a mammogram, without taking the breast tissue into account. Briefly, a learning stage is firstly developed in order to acquire the knowledge of the shape and the size of mammographic masses. Following, the algorithm looks for masses in a new mammogram following a template matching scheme. Results are evaluated over different databases and are given using Receiver Operating Characteristic (ROC) and Free-Response Receiver Operating Characteristic (FROC) analysis. The conclusion is that the method has large accuracy but with the penalty of obtaining a number of false positives.

Thus, in Chapter  a proposal for false positive reduction based on the statistical analysis of the RoIs is elucidated and extensively evaluated using different databases and ROC and FROC analysis. The evaluation is done in two steps: firstly the algorithm is evaluated using a set of manually segmented RoIs, while secondly the overall performance of the system (Bayesian template matching plus false positive reduction) is analyzed.

Chapter  describes how to introduce the breast tissue information into both algorithms. FROC and ROC analysis demonstrates the usefulness of such information. Finally, the thesis concludes with Chapter , where the conclusions and the ways in which further work is directed are covered. Moreover, in this last chapter, a list of the publications related to this thesis is included.

Moreover, three appendixes are given. The first one is related to the breast profile segmentation. Firstly, the proposal used in this work for segment the mammograms is explained. Secondly, a new algorithm which better adjusts to the external boundary of the mammogram is presented. The second appendix is focused on explaining the main characteristics of the databases of mammograms used in this work, which are four different databases in total. The last appendix explains the strategies used to evaluate the proposals.