ELECTRIC: AI-powered CT reconstruction slashes errors 70%

A new arXiv paper from Ge Wang introduces ELECTRIC, an evidential learning-enhanced CT reconstruction framework that closes the loop between deep learning and physics-based iterative reconstruction. The method uses a Normal-Inverse-Gamma evidential network to generate both an image proposal and an error-predictive epistemic-uncertainty surrogate. This uncertainty is converted into an adaptive precision field, which is then injected into a Poisson-weighted maximum-a-posteriori (MAP) update. The result is a self-correcting loop that treats prior confidence as a learned state variable, enabling the model to refine its reconstructions iteratively without losing measurement fidelity.

In simulation studies on the AAPM Mayo Clinic Low-Dose CT dataset, ELECTRIC achieved roughly 70% lower reconstruction error than filtered back-projection on held-out patients. The epistemic uncertainty estimates proved to be error-predictive, meaning the model can flag its own mistakes and support selective trust in clinical workflows. The adaptive precision reconstruction also matched or beat a validation-tuned fixed prior and remained far more robust when prior strength was misspecified. These results validate the full closed-loop pipeline, while the paper points to formal uncertainty calibration and joint training as the main next steps. For researchers and developers in medical imaging, ELECTRIC offers a concrete template for combining evidential deep learning with physics-guided reconstruction to improve diagnostic image quality at lower radiation doses.