Show abstract - PDF - Reviews

Ultrasound and radar signals are useful for medical imaging due to their non-invasive, non-radiative, low-cost, and accessible nature. However, traditional imaging techniques lack resolution, contrast, and physical interpretation. Quantitative medical imaging is necessary for this purpose, as it provides a visual representation of physical characteristics. However, current techniques have drawbacks, including convergence to local minima and delayed results, which can lead to unsatisfactory outcomes. To address these limitations, we propose a neural network that incorporates the symmetries and properties of the received signals to achieve real-time quantitative mappings of physical properties. Our method achieves high accuracy using several numerical metrics for complex shapes with less than 0.15 seconds per test sample, compared to 0.75-2 hours for the competing method.
Hide abstract