Spectral characterization of mammographic tissue for computer aided diagnosis of malignant masses

An approach for the analysis of the spectral properties of digitized mammograms for computer aided diagnosis is presented. The approach is developed using 206 regions of interest (ROIs) extracted from 103 normal and 103 malignant mass cases selected from the Digital Database for Screening Mammography (DDSM) available from the University of South Florida. A spectral definition is proposed in terms of a linear function of the local slope of the modified, circularly averaged periodogram across the entire spectrum. The local slope is estimated in a least squares sense for each point in the spectrum as a function of local neighboring samples. The proposed spectral definition is evaluated for the discrimination of malignant versus normal ROIs. Results are summarized by receiver operating characteristic (ROC) curve analysis. For the cases studied, maximum detection performance is achieved with an area under the ROC curve (AUC) of 0.9223 and a partial AUC at 90% sensitivity of 0.712. These results suggest that the proposed spectral signature is useful as a computationally fast and effective approach for the characterization of malignant masses in digitized mammograms.