An Integration Between SVM Classifiers and Multi-Resolution Techniques for Early Breast Cancer Detection

Because of the high contrast between the dielectric properties of normal and malignant breast tissues at microwave frequencies, microwave imaging techniques seem to be very attractive diagnosis methods for cancer detection [1][2]. In such a framework, inverse scattering methods are very promising tools, but their practical application is strongly limited by the need of 3D reconstructions, high spatial resolutions, and fast processing. Recently, to reduce the high computational costs and to fit the real‐time requirements, inversion methods based on learning by example techniques have been proposed [3]. LBE approaches based on support vector machines (SVMs) [3] and neural networks (NNs) [4] have been satisfactorily applied in various and complex electromagnetic problems. When dealing with breast cancer detection, the inversion process is recast as a classification or regression problem where the unknowns are retrieved from the data (i.e., the electric field samples collected in an external observation domain) by approximating the unknown relation data‐unknowns through an off‐line data fitting procedure (training phase). Once the training procedure (performed once and off‐line) is completed, the characteristics of the malignant breast tissue are real‐time estimated in the testing phase. In such a work, the detection problem is addressed by integrating a SVM‐based classifier with an iterative multi‐zooming procedure. More in detail, a succession of approximations of a probability map of the presence of pathology is determined. At each step, the spatial resolution of the risk‐map is improved in a limited set of regions of interest (ROIs) defined at the previous zooming step and characterized by a greater value of the occurrence probability of a malignant tissue. The multi‐step procedure is stopped when a stationary condition on the probability and on the number of ROIs is reached. The achievable trade‐off between computational complexity and spatial resolution is preliminary assessed by discussing a selected set of numerical simulations concerned with both noiseless as well as corrupted data.