Electromagnetic Breast Imaging and Uncertainty Quantification with Bayesian Neural Networks

Different scenarios of breast imaging segmentation and estimation of its electromagnetic (EM) or/and ultrasonic (US) parameters are presented, part of a two-pronged dynamic using Bayesian Neural Networks, the first being to continue obtaining more and more qualitative results, while the second is to propose a precise and detailed framework for the choices made on the basis of the underlying physics and/or empirical behavioral observations of the a priori distributions of Bayesian neural network parameters. We propose a methodological participation applicable to Bayesian convolutional neural networks, showing the efficiency of such an approach with simulated and real EM or/and US breast imaging datasets. A Bayesian data fusion framework (e.g.multi-frequency data and multi-physics data) is hence introduced since EM low-resolution and US high-resolution imaging complementarity has been shown to improve the quality of the reconstructed image, providing well-contrasted zones.