Unlocking on-Board Dual-Energy: Fast & Robust Material Decomposition for the Kv Dual-Layer Imager
Abstract
Purpose
Conventional dual-energy material decomposition methods suffer from significant noise amplification, which limits their clinical utility for tasks such as intra-fraction markerless tumor tracking. This study introduces PRISM, an open-source Julia framework specifically designed for real-time, iterative, regularized dual-energy material decomposition on a novel prototype of kV dual-layer imager (DLI).
Methods
Dual-energy projections of a TOR-18FG phantom and of a patient thorax were acquired using the DLI. We implemented a penalized weighted least-squares objective function for iterative material decomposition. Four regularization strategies were investigated: quadratic, edge-weighted quadratic, non-local similarity, and a proposed new cross-similarity regularization, which enforces consistency across both spatial and spectral domains. Decomposition performance into water and bone components was evaluated using the TOR-18FG phantom, quantifying the trade-off between noise, signal bias, and spatial resolution. The methods were validated on the patient thoracic projections. Computation times were compared between a CPU sparse-matrix implementation and a custom GPU matrix-free implementation.
Results
The proposed cross-similarity regularization achieved up to a fivefold noise reduction in water-equivalent images while preserving spatial resolution, outperforming local quadratic methods. Unlike standard similarity regularization, which induced signal bias (up to 19%) when using larger search windows, our new cross-similarity maintained bias below 1%. In patient studies, cross-similarity enhanced soft-tissue visibility and preserved fine lung structures more effectively than local methods. The GPU-accelerated implementation achieved decomposition times under 50 ms, approximately two orders of magnitude faster than the CPU version.
Conclusion
PRISM provides a robust framework for high-quality dual-energy material decomposition on the DLI. The novel cross-similarity regularization offers a high noise reduction level while preserving spatial resolution and minimizing bias. The GPU implementation achieves real-time processing speeds, paving the way for clinical applications such as intra-fraction markerless tumor tracking.