On Anatomical Prior Kernel Distortion for Tomographic Microwave Imaging Stroke Monitoring

This paper investigates the role of prior internal structural, morphological, and dielectric information in constructing linearized inversion kernels for microwave imaging-based brain stroke monitoring, with emphasis on structural-model mismatch and regularization effects. Although detailed anatomical models are often assumed to improve reconstruction accuracy, their availability, reliability, and practical integration remain limited in realistic clinical scenarios. A systematic methodological framework is therefore adopted to assess how the inclusion, simplification, or absence of morphological information affects the performance of the inversion kernels constructed under the Born approximation and regularized using the truncated singular value decomposition. A multi-patient numerical study is used to capture anatomical variability and evaluate inversion operators with progressively reduced structural complexity, ranging from multi-tissue representations to homogeneous models. Stroke evolution, including both hemorrhagic and ischemic cases, is simulated and analyzed under varying spatial-resolution, lesion-morphology, and lesion-location conditions. In addition, this work first harmonizes the terminology and definitions of imaging metrics commonly used in recent microwave medical imaging literature, providing a consistent framework for their interpretation and cross-study comparison, and second proposes a unified statistically based metric, referred to as the Spearman-weighted metric, providing a global score that enables direct and consistent comparison across inversion kernels. The results indicate that reconstruction quality is only weakly dependent on the level of structural detail embedded in the inversion kernel. In many cases, simplified models are sufficient to capture the main features of stroke evolution, suggesting that highly detailed patient-specific information may not be strictly necessary for monitoring-oriented applications. Overall, the proposed framework provides a general methodological basis for evaluating structural-model mismatch in inversion kernels and supports the use of reduced-complexity inversion strategies when accurate prior information is unavailable.