Clutter Removal in the Latent Space Using Data-Driven Method for Clinical Microwave Head Imaging

For reliable microwave medical imaging, clutter signals including scattering from healthy tissues and interference in the imaging system should be mitigated. To achieve this objective, a data-driven clutter removal method is proposed. Since clutter occupies the most signal energy in medical microwave imaging, it contributes to the primary variance of signals collected from different objects. Therefore, clutter features can be found in a latent space defined by the clutter directions of the scattered signals. Clutter directions are constructed using small-sized simulated, lab experiments or healthy volunteers’ datasets. Subsequently, clutter signals are extracted by projecting the features in the latent space back to the original signal space, creating clutter-free signals, which are processed using optimization-based confocal imaging. The method is validated through simulations, lab experiments, and a clinical trial using a 16-antenna imaging system. In simulation and experiment studies, numerical head models (458 cases) and fabricated multilayer head phantoms (488 cases) are used to build datasets, in which 23-44% of cases are used to derive the clutter directions, while the remaining cases are used in the assessment. In the clinical trial, signals from 43 healthy volunteers were used to form the latent space and calculate the clutter directions of signals. The method was used in a pilot clinical trial to generate brain images for 38 stroke patients. Notably, our approach achieves detection accuracies of 90%, 86%, and 81% in simulation, lab experiments, and pilot clinical trials, respectively, outperforming other methods, which have accuracies below 42%.