A method for characterizing tumoral masses in mammograms is described.
It is based on two significant features of breast cancer, namely the
degree of spiculation and the presence of fuzzy areas in the boundary
of the tumor. The quantization of these features is performed by means
of polar and pseudopolar representations of the neighborhood of the t
umor to take advantage of its symmetry. The degree of spiculation is d
educed from a shape parameter that characterizes the irregularity of t
he boundary after enhancing divergent structures. The radial component
of the gradient on the boundary provides a measurement of fuzzy appea
rance. In the next step, the identification of malignant and benign tu
mors is achieved with a two-level hierarchical classification using a
Bayesian classifier. A set of 19 masses are processed and results are
presented and discussed. (C) 1996 Society of Photo-Optical Instrumenta
tion Engineers.