A generalized supervised training approach for adaptive spectral harmonization in photon-counting CT

Photon-counting CT (PCCT) has the potential to provide superior spectral separation and improve material decomposition accuracy. However, PCDs face challenges of pixel-level inhomogeneity and instability, necessitating frequent spectral calibration and specialized processing adaptive to different spectral responses. In this work, we introduce a generalized supervised training method for adaptive spectral harmonization in PCCT. A multi-layer perceptron (MLP) network is trained to achieve material decomposition regardless of individual PCD pixel responses, resulting in artifact-free images and accurate material quantification, as validated in simulations and experiments.