A Pruned Neural Network for Microwave Imaging

Inverse scattering problems arise in several fields and can be solved using linear approximations or by accounting fully for their intrinsic nonlinearity. In recent years, neural networks (NNs) have attracted considerable interest in solving inverse problems due to their ability to approximate complex mappings between sets. However, designing an effective NN architecture and defining a suitable training strategy remain challenging tasks. In this work, we investigate the potential of a NN for reconstructing two-dimensional metallic scatterers in a multi-monostatic configuration. Particular attention is devoted to the dimensioning of the NN, which is optimized using a pruning strategy aimed at reducing model complexity. Inspired by biological synaptic pruning, the proposed approach removes weakly connected neurons after training, yielding a more compact architecture without degrading reconstruction accuracy. The performance of the NN-based inversion is assessed through a comparison with a linear approximation of the inverse scattering operator, demonstrating the effectiveness of the proposed methodology. The paper also includes comparisons with a one-dimensional U-Net and with the contrast source inversion (CSI) method, providing nonlinear benchmarks for the proposed approach. A sensitivity analysis with respect to different neuron-reduction percentages is carried out to quantify the trade-off between model compactness and reconstruction accuracy. The robustness of the method is further assessed through a stability analysis over multiple noise realizations and through experimental tests involving off-centered scatterers. Finally, a preliminary extension to a three-dimensional scattering scenario is presented, in which convolutional neural networks (CNNs) are employed to exploit the spatial structure of the two-dimensional support.