ELECTRIC: Evidential Learning-Enhanced CT Reconstruction via Iterative Correction

Abstract

We introduce ELECTRIC (Evidential Learning-Enhanced CT Reconstruction via Iterative Correction), a physics-guided Bayesian framework for computed tomography (CT) reconstruction. The method employs an evidential neural network to generate both an image proposal and an error-predictive epistemic-uncertainty surrogate. This surrogate is converted into an adaptive precision field and integrated into a Poisson-weighted Maximum A Posteriori (MAP) update. The resulting image–evidence–precision–reconstruction loop treats prior confidence as a learned state variable within iterative reconstruction. Beyond the theoretical formulation, we present two simulation studies using image slices from the AAPM Mayo Clinic Low-Dose CT dataset: a mechanism-validation pilot with transparent surrogate estimators, and a feasibility study where a trained Normal-Inverse-Gamma evidential network drives the full closed loop. On held-out patient data, the learned prior mean reduces reconstruction error by approximately 70% relative to filtered back-projection (FBP). The learned epistemic uncertainty proves error-predictive, enabling selective trust in reconstructions. The physics-guided update restores measurement consistency, while the adaptive-precision reconstruction matches or exceeds a validation-tuned fixed prior and demonstrates markedly greater robustness to prior-strength misspecification. These results validate the complete ELECTRIC closed-loop pipeline and highlight formal uncertainty calibration and joint training as key directions for future work.

1. Introduction

Computed tomography (CT) reconstruction has undergone significant evolution in recent years, transitioning from purely analytic methods like filtered back-projection (FBP) to iterative approaches and, more recently, deep learning-based techniques. As of 2026, the field increasingly recognizes the importance of integrating physics-based constraints with learned representations to improve both image quality and reliability.

Traditional deep learning approaches to CT reconstruction often operate as black boxes, providing point estimates without reliable uncertainty quantification. This limitation is particularly critical in medical imaging, where clinical decisions depend on trustworthy confidence assessments. Bayesian formulations offer a principled solution, but full posterior inference remains computationally intractable for high-dimensional imaging problems.

2. Methodology

2.1 Evidential Learning Framework

The ELECTRIC framework leverages evidential deep learning, a paradigm that treats neural-network outputs as parameters of a higher-order probability distribution. Specifically, we employ a Normal-Inverse-Gamma (NIG) evidential network that predicts both the image estimate and an associated epistemic-uncertainty surrogate. Unlike conventional Bayesian neural networks, this approach enables closed-form uncertainty estimation without sampling or variational approximations.

2.2 Physics-Guided Iterative Correction

The core innovation of ELECTRIC lies in its closed-loop architecture that integrates evidential predictions directly into the reconstruction process. The pipeline proceeds as follows:

1. An evidential network proposes an initial image estimate alongside its epistemic uncertainty map
2. The uncertainty surrogate is converted into an adaptive precision field
3. This precision field is incorporated into a Poisson-weighted MAP update that enforces measurement consistency
4. The updated reconstruction feeds back into the evidential network for further refinement



This image–evidence–precision–reconstruction loop fundamentally reinterprets prior confidence as a learned state variable, allowing the system to adaptively balance physical fidelity and learned priors throughout the iterative process.




2.3 Theoretical Justification




We provide theoretical analysis demonstrating that the ELECTRIC formulation maintains the convergence properties of standard iterative reconstruction while introducing adaptive regularization informed by evidential uncertainty. The precision field derived from the uncertainty surrogate naturally downweights uncertain regions during physics updates, preventing the learned prior from overriding measurement information in areas where confidence is low.




3. Experiments and Results




3.1 Experimental Setup




We conducted two simulation studies using image slices from the AAPM Mayo Clinic Low-Dose CT dataset. The first pilot study evaluated mechanism validation using transparent surrogate estimators, establishing the correctness of the uncertainty-informed update strategy. The second feasibility study employed fully trained NIG evidential networks to drive the complete closed loop across the entire reconstruction pipeline.




3.2 Key Findings




On held-out patient data, ELECTRIC achieves substantial improvements over conventional baselines:



• Reconstruction Accuracy: The learned prior mean reduces reconstruction error by approximately 70% compared to filtered back-projection.
• Uncertainty Quality: The learned epistemic uncertainty is genuinely error-predictive, enabling selective trust mechanisms where clinicians can identify high-confidence reconstruction regions.
• Measurement Consistency: The physics-guided update step successfully restores consistency with the measured projection data, confirming that the learned prior does not override physical constraints.
• Robustness: The adaptive-precision reconstruction matches or exceeds a validation-tuned fixed prior while remaining notably more robust to misspecification of the prior strength parameter—a critical practical advantage given the difficulty of tuning such parameters in clinical settings.



4. Discussion and Future Directions




While the complete ELECTRIC closed-loop pipeline demonstrates strong performance, our analysis identifies two principal avenues for future work:



1. Formal Uncertainty Calibration: Although the epistemic uncertainty is error-predictive, formal calibration of the uncertainty estimates remains necessary to ensure they reflect true confidence levels suitable for clinical decision-making.


1. Joint Training: End-to-end training of the evidential network together with the iterative reconstruction process offers potential for further performance gains. Current results were obtained with separately trained components; joint optimization could better align the learned priors with the physics-guided update dynamics.



These directions align with the broader 2026 trend toward trustworthy AI in medical imaging, where uncertainty-aware architectures that maintain physical fidelity are increasingly recognized as essential for clinical translation.




5. Conclusion




ELECTRIC represents a significant advance in physics-guided deep learning for CT reconstruction, demonstrating that evidential uncertainty can serve as a practical mechanism for adaptive regularization within iterative algorithms. The substantial error reduction, reliable uncertainty estimation, and robustness to hyperparameter misspecification suggest that this approach has strong potential for clinical deployment in low-dose CT settings, where both image quality and confidence assessment are paramount.

via ArXiv CV