Clutter Removal for Microwave Head Imaging via Self-Supervised Deep Learning Techniques

Microwave head imaging is challenging due to the dominance of clutter signals caused by the strong reflections at the boundary of the head and skull in addition to the heterogeneous nature of the head tissues. These clutter signals complicate the detection of anomalies like strokes and make both traditional and deep-learning-based imaging algorithms less effective. For example, to adapt to different environments, extensive tuning is required for traditional algorithms, while a huge amount of data is needed to train deep-learning models. To this end, a novel deep-learning-based clutter removal approach in microwave head imaging is proposed. The proposed deep learning model is self-supervised and unpaired, and can thus utilize much larger amounts of data, which would otherwise be prohibitively difficult to collect. The model includes two generators to learn the mapping function from mixed signals and the target signal alone to remove clutter and ensure producing target signals that match the original mixed signals. To achieve self-supervised learning, two discriminators are used for judging the predictions from both generators by comparing the predictions with the real signals. Using the peak signal-to-noise ratio and the structural similarity index measure, the experimental results using a 16-antenna head imaging system operating across the band 0.5–2 GHz confirm that the presented solution outperforms existing methods in removing clutter and enabling accurate target localization. The proposed solution is adaptable and scalable and can thus be generalized to other domains.