We propose developing statistical fractal features which quantify lesion border roughness on mammograms and using these features to distinguish malignant and benign breast lesions. Objective measures of lesion roughness are important in the diagnosis and staging of breast cancer. In this novel approach, we generate a space of fractal models, and evaluate the fractal dimension (fd) of the lesion from the statistics of the model space. Using intensity as the third dimension, we generate a 3-dimensional image of the mammogram and select a set of intensity levels found in the lesion. For each selected intensity level, we construct a binary thresholded image and trace the lesion border. Analysis is restricted to segments of thresholded borders which are in high-gradient portions of the 3-dimensional image and can be determined with high precision. Each selected boundary segments is analyzed to identify self-affine subsegments which are modeled using multiple fractal interpolation functions having known fd values. Thus we construct a large sample space of fd values which are computed from high-precision boundary traces occurring on a range of threshold levels. The statistics of the fd sample space are the features used to distinguish between malignant and benign lesions. PROPOSED COMMERCIAL APPLICATION Our product will have significant value to the diagnostician who must distinguish malignant from benign breast lesions. The algorithm is readily integrated into both computer-aided diagnosis systems and digital mammogram systems which display and process mammographic images for the expert diagnostician
