A research group, helmed by Dr. Liu Dong from the Spintronics and Magnetic Resonance Laboratory at the University of Science and Technology of China (USTC), has made a substantial leap forward in the interdisciplinary realm of computational imaging and artificial intelligence. They have introduced a self-supervised learning framework named Physics-Informed Neural Network Compensation (PhyNC), which effectively tackles the central issue of uneven sensitivity distribution in medical Electrical Impedance Tomography (EIT). Their findings have been published in the esteemed journal IEEE Transactions on Pattern Analysis and Machine Intelligence, a top-tier publication in the AI domain.

EIT technology reconstructs internal impedance distributions by reversing surface currents and boundary voltages, boasting the benefits of being non-invasive, real-time, dynamic, and safe. Nevertheless, influenced by the soft-field effect of currents, its measurement sensitivity drops notably with depth. This results in deep-region information being readily overshadowed by noise, rendering the image reconstruction inverse problem highly ill-posed. To overcome this hurdle, the research team delved deeply into the physical mechanisms underlying EIT and developed a sensitivity-aware mechanism. By employing hierarchical mapping based on physical priors, the neural network can discern the spatial distribution of sensitivity in the physical field, much like the human eye's 'gaze' mechanism. It automatically allocates more representational power to compensate in low-sensitivity areas while imposing constraints to minimize noise in high-sensitivity regions.

Furthermore, the team introduced a hybrid representation technique that combines multi-scale embeddings with Fourier feature projections, along with a self-devised frequency regularization strategy. This method significantly boosts the neural network's reconstruction prowess and robustness in both high- and low-sensitivity regions. Without the need for labeled data, the framework achieves high-fidelity and robust image reconstruction in both simulated and real-world experiments. Particularly in low-contrast, low-sensitivity central areas, it can accurately reconstruct geometric shapes, showcasing remarkable noise resistance and generalization capabilities across various mesh resolutions.

This accomplishment not only presents a novel 'neural compensation' approach to the challenge of non-uniform sensitivity in image reconstruction inverse problems but also establishes a robust foundation for the practical deployment of EIT technology in portable medical monitoring, flexible electronic skin, and industrial non-destructive testing.