Increasing the Reliability of the Proposal

This kind of further work is directed to increase the performance of the algorithm. In this sense, a set of different directions are possible: to improve the false positive reduction approach, to reduce the computational time, to change the initial breast profile segmentation algorithm, or also to use more than one mammographic view in order to increase the performance of the full algorithm.

As we have shown, the performance of the false positive reduction algorithm is database dependent. Possible further work is directed to reduce this dependency, which could be done applying the standard mammogram form (also known as the $h_{int}$ approach) proposed by Highnam and Brady [73]. This approach relies on a detailed knowledge of the mammographic systems and the imaging parameters, and as such might be less appropriate for mammograms where this information is not available (see also Blot and Zwiggelaar [15] and Highnam et al. [74] for a detailed discussion). In these later cases, only the study of common features in both databases will provide information for the correct normalization of both databases.

The reduction of the computational cost is necessary if we aim at obtaining an online tool, otherwise this is less important because we can execute the algorithm as a batch process which can be finished before the experts read the images. We are almost sure that translating the algorithm to a more efficient programming environment like C++ (instead of using Matlab) will improve the overall computational cost. On the other hand, a more difficult and also more interesting problem, is the fact that if we want to add new representative cases in the training database all models need to be recomputed. Thus, additional further work will focus on the development of an incremental training step.

The third mentioned further direction of this group is to apply the new proposal of breast profile because the one used in this work removes some pixels near the skin-line (see Appendix  for a detailed discussion). However, the use of this algorithm will not represent a great change to the obtained results of the mass detection algorithm, although the performance of the breast tissue classifier algorithm could be greatly modified, because we are sure that those pixels represent a new cluster for the Fuzzy C-Means algorithm. Thus, probably, we would have to use three classes instead of the two used in this thesis, which will represent: fatty tissue, dense tissue, and pixels near the skin-line. If this is the case when segmenting the images, only information coming from dense and fatty classes will be useful, and thus, the rest of the classifier algorithm (features used and classifier combination) will still be useful.

Finally, we would like to improve the performance of the algorithm using information coming from the analysis of other mammographic views, either from the same breast or from the complementary one. In this sense, while the CC-MLO registration needs a major study, we can used different already developed algorithms [114,] to obtain the MLO-MLO registration.