One-Shot Physics-Informed Implicit Neural Representations for Low-Dose Dynamic Myocardial Perfusion CT Reconstruction
Abstract
Purpose
Dynamic myocardial perfusion CT (MPCT) serves as a critical tool for functional assessment of multiple diseases, including coronary stenosis and myocardial ischemia. However, the continuous acquisition required to track contrast kinetics incurs substantial radiation dose to patients. Low-mAs protocols offer a dose-reduction strategy but suffer from severe quantum noise and streak artifacts, which degrade spatiotemporal fidelity and prevent the accurate derivation of quantitative perfusion maps. We propose a robust machine learning method, MPCT-INR, to recover anatomical structures and dynamic contrast information from low-dose projection data.
Methods
To exploit inter-frame correlations in contrast dynamics, we use an Implicit Neural Representation (INR) to tightly modelthe spatiotemporal dynamic CT sequence. A physics-informed, two-stage optimization is adopted to recover the dynamic sequence within a ‘one-shot’ learning framework. Firstly, the network is pre-trained using noise-corrupted, per-framefiltered back-projection (FBP) reconstructions to provide a stable initialization. Secondly, the network is jointly optimized with a composite loss function consisting of a data fidelity term (minimizing the discrepancy between forward-projected dynamic CT images queried from the spatiotemporal INR and the measured projection data), and regularization terms that enforce spatial smoothness and temporal continuity.
Results
Validated on low-dose XCAT simulations, the proposed MPCT-INR demonstrated superior noise suppression and artifact removal. Quantitatively, it achieved a 20.1% improvement in PSNR and a 58.1% reduction in relative error compared to the FBP baseline. Notably, compared to an explicit tensor-based self-supervised method, which suffered from semi-convergence and noise overfitting, the MPCT-INR method exhibited robust stability during optimization. Furthermore, the reconstructed time-attenuation curves for ventricles and myocardium showed high temporal fidelity, accurately capturing wash-in and wash-out dynamics consistent with the ground truth.
Conclusion
MPCT-INR effectively reconstructs anatomical structures and recovers dynamic contrast variations from low-dose data, outperforming baseline methods and demonstrating great promise for dose-efficient dynamic myocardial perfusion imaging in clinical applications.