Uncertainty-Aware Conductivity Reconstruction via a 2.5D Patch-to-Patch Data-driven Model

Motivation: To address the lack of confidence in end-to-end data-driven conductivity reconstructions. Goal(s): To propose a method for calculating voxel-wise variability of reconstructed conductivity maps and complex B1+ field data discrepancy using a feedback loop, facilitating a comprehensive assessment of uncertainty in deep-learning electrical-property-reconstructions. Approach: We utilize an end-to-end model to derive conductivity and voxel-wise variability maps, from which complex B1+ maps are predicted using a finite difference approach as data discrepancy index. Results: The proposed method show high-quality conductivity maps. The voxel-wise variability and data discrepancy index provide confidence on the reconstructed conductivity maps. Impact: A 2.5D uncertainty-aware data-driven framework is developed for conductivity reconstructions. This approach enhances estimation accuracy and quantifies variability, providing insights into model reliability and improving performance compared to conventional physics-based and end-to-end deep-learning methods.