Early breast anomalies detection with microwave and ultrasound modalities

Imaging of the breast for early detec-tion of tumors is studied by associating microwave (MW) and ultrasound (US) data. No registration is enforced since a free pending breast is tackled. A 1st approach uses prior information on tissue boundaries yielded from US reflection data. Regularization incorporates that two neighboring pixels should exhibit similar MW properties when not on a boundary while a jump allowed otherwise. This is enforced in the distorted Born iterative and the contrast source inversion methods. A 2nd approach involves deterministic edge preserving regularization via auxiliary variables indicating if a pixel is on an edge or not, edge markers being shared by MW and US parameters. Those are jointly optimized from the last parameter profiles and guide the next optimization as regularization term coefficients. Alternate minimization is to update US contrast, edge markers and MW contrast. A 3rd approach involves convolutional neural networks. Estimated contrast current and scattered field are the inputs. A multi-stream structure is employed to feed MW and US data. The network outputs the maps of MW and US parameters to perform real-time. Apart from the regression task, a multi-task learning strategy is used with a classifier that associates each pixel to a tissue type to yield a segmentation image. Weighted loss assigns a higher penalty to pixels in tumors when wrongly classified. A 4th approach involves a Bayesian formalism where the joint posterior distribution is obtained via Bayes’ rule; this true distribution is then approximated by a free-form separable law for each set of unknowns to get the estimate sought. All those solution methods are illustrated and compared from a wealth of simulated data on simple synthetic models and on 2D cross-sections of anatomically-realistic MRI-derived numerical breast phantoms in which small artificial tumors are inserted.