A two-stage workflow for detecting and monitoring tumors in the human breast with an inverse scattering-based technique is presented. Stage 1 involves a phaseless bulk-parameter inference neural network that recovers the geometry and permittivity of the breast fibroglandular region. The bulk parameters are used for calibration and as prior information for Stage 2, a full phase contrast source inversion of the measurement data, to detect regions of high relative complex-valued permittivity in the breast based on an assumed known overall tissue geometry. We demonstrate the ability of the workflow to recover the geometry and bulk permittivity of the different sized fibroglandular regions, and to detect and localize tumors of various sizes and locations within the breast model. Preliminary results show promise for a synthetically trained Stage 1 network to be applied to experimental data and provide quality prior information in practical imaging situations.

We present a neural network architecture to determine the volume and complex permittivity of grain stored in metal bins. The neural networks output the grain height, cone angle and complex permittivity of the grain, using the input of experimental field data (S-parameters) from an electromagnetic imaging system consisting of 24 transceivers installed in the bin. Key for practical applications, the neural networks are trained on synthetic data sets but generate the parametric information using experimental data as input, without the use of calibration objects or open-short-load measurements. To accomplish this, we formulate a data normalization scheme that enables the use of a loss function that directly compares measured S-parameters and simulation model fields. The normalization strategy and the ability to train on synthetic data means we do not need to collect experimental training data. We demonstrate the applicability of this synthetically trained neural network to experimental data from two different bin geometries, and discuss the ability of these neural networks to successfully infer parameters that can be used for grain inventory management. Our neural-network-based approach enables rapid inference, providing a more cost-effective long-term solution than existing optimization-based parametric inversion methods.

No abstract

We demonstrate the recovery of simple geometric and permittivity information of breast models in an experimental microwave breast imaging system using a synthetically trained machine learning workflow. The recovered information consists of simple models of adipose and fibroglandular regions. The machine learning model is trained on a labelled synthetic dataset constructed over a range of possible adipose and fibroglandular regions and the trained neural network predicts the geometry and average permittivty of the adipose and fibroglandular regions from calibrated experimental data. The proposed workflow is tested on two different experimental models of the human breast. The first model is comprised of two simple, symmetric phantoms representing the adipose and fibroglandular regions of the breast that match the model used to train the neural network. The second, more realistic model replaces the symmetric fibroglandular phantom with an irregularly shaped, MRI-derived fibroglandular phantom. We demonstrate the ability of the machine learning workflow to accurately recover geometry and complex valued average permittivity of the fibroglandular region for the simple case, and to predict a symmetric convex hull that is a reasonable approximation to the proportions of the MRI-derived fibroglandular phantom.

No abstract