An analysis of perception errors through image structure

Breast cancer is one of the leading causes of death among American women. Currently mammography is the most widely used technique to screen for this disease. Nonetheless, 10–30% of all cancers in the breast go unnoticed by the interpreting radiologist, being only found retrospectively, once the cancer has grown. In this thesis, we will study the areas in mammograms that yield correct decisions, being these True Positives or True Negatives, and the ones that yield incorrect decisions, namely, False Positives or False Negatives.; We showed that each decision outcome has a characteristic signature in the spatial frequency domain, and this spatial signature can be recognized by a pattern recognition system. Furthermore, we showed that individual observers react in different ways to the same elements in the image, and it is possible to predict how each observer will report a given area on a mammogram, by using an artificial neural network which receives as input the energy values of the decomposition of the image patch using a wavelet packets tree two-levels deep. We also showed that this decomposition permits the separation of the image patch between truly normal tissue and tissue where the cancer, being deemed visible or not, later grew.