Local orientation distribution as a function of spatial scale for detection of masses in mammograms

A method has been developed for detection masses in mammograms by analysis of local orientation patterns. Concentration of gradient and line orientation computed at a fine scale reveals the presence of masses and spiculation, respectively. In this paper a new computational approach is presented which allows efficient computation of these features as a continuous function of spatial scale. It is shown that by using these scale signatures estimates of mass size can be readily obtained. Experimentally it was found that mass size estimates can be used to improve mass detection, while full exploitation of the information represented by the scale signatures is expected lead to further improvement. Results are presented for detection of malign masses in a database of 264 mammograms representing 71 consecutive cancers found in screening.