Contributions

Thus, we consider that the main contributions of this Thesis are:

• An extensive survey of mass segmentation algorithms, which are classified in both the strategy and the features used to segment the mammograms. Moreover, from the quantitative comparison of eight of these methods we prove that the breast density and the size and shape of the masses are three parameters which significantly influence the performance of the algorithms.

• A survey on breast classification methods. We review the main methods found to classify the breasts according to their internal density, highlighting their strategy, features, and the classification used. Moreover, from the quantitative comparison of the strategies, we found that grouping according to the pixel appearance outperforms current strategies.

• A new algorithm for breast density classification based on tissue segmentation and a posterior ensemble classification algorithms, which is exhaustively evaluated using MIAS and DDSM database.

• A new proposal on mass detection, which takes the three parameters above mentioned into account: the breast density which is known a priori, and the shape and size of the masses which are known during the matching step using the proposed Bayesian algorithm. The performance is tested using MIAS and Trueta databases while DDSM is used for training.

• A new false positive reduction algorithm based on the 2DPCA approach. We have shown that this algorithm performs better when training and testing with the same database.

• Finally, as a result of the close relationship with the Radiologic Department of Hospital Josep Trueta of Girona, a new full-field mammographic database has been compiled and made available.