Mitigating Mesh-Induced Errors in Time-Domain Simulation of Medical Microwave Imaging

With the rapid development of medical microwave imaging systems, their accurate simulation becomes important to ensure the reliability of the developed hardware and software before moving to clinical tests. This is particularly important when using machine learning techniques that require a large volume of simulated data for training. Since the target responses in medical microwave imaging are quite weak, the numerical errors and variations caused by different meshing used in numerical simulations most likely mask the target response; thus, interfering with the pattern of the simulated dataset and resulting in misleading results. Since most available bio-models are voxel-based and can only be meshed using hexahedra, this article focuses on time-domain solvers that use hexahedral meshing. It investigates numerical modeling errors caused by mesh variations in realistic three-dimensional (3D) simulations, their impact on data distribution for machine learning algorithms, and the trade-off between convergence and total simulation time. A realistic 3D microwave head imaging system is used as an example. A delta-simulation approach is presented to eliminate data variations in the constructed dataset caused by inconsistent mesh distributions across simulated cases. The kernel principal component analysis and k-means clustering techniques are used to evaluate the proposed approach. The assessments show that the proposed delta-simulation method can generate a dataset with a more informative data distribution, thereby facilitating subsequent machine learning algorithms. In addition, the delta-simulation balances accuracy and efficiency by maintaining an acceptable number of mesh elements that can be simulated in a reasonable time.