Abstract Objective: Spectral CT data from Photon-Counting CT (PCCT) enables material decomposition. Mechanistic approaches such as Maximum Likelihood Estimation (MLE) are noise sensitive. Deep learning alternatives mitigate this issue, but their accuracy remains limited due to lack of incorporation of underlying physics principles and lack of ground truth data. This study aims to develop and validate a physics-informed deep-learning model, trained on validated simulated data, to decompose spectral CT images into density\ () and effective atomic number (Zeff) maps. Methods: The training dataset included simulated abdominal PCCT scans from 32 human models with corresponding ground truth. The scans were obtained at two clinical dose levels, four detector energy thresholds, different iodinated contrast agent concentrations and reconstructed using three clinically-used kernels. A Generative Adversarial Network (GAN) was trained with and without a physics-informed regularization loss to estimate ρ and Zeff maps. Model performance was evaluated on 16 computational phantoms and validated on 6 clinical cases. A reader study was performed on 30 image slices to assess the comparative performance of ρ and Zeff maps to multi-rendered Virtual Monochromatic Images (VMIs) for assessing liver lesion conspicuity. Main Results: With physics-informed regularization, NRMSE of 1. 29% and 0. 68%, SSIM of 0. 99 and 0. 99, and PSNR of 29. 8dB and 29. 04dB were achieved. A maximum RMSE of 5. 45% was achieved on clinical data. Reader study results showed ρ and Zeff images had higher conspicuity scores compared to VMIs (median: 4. 52 vs. 4. 13; 95% CIs: 4. 19, 4. 52 vs. 4. 01, 4. 31). The study showed equivalent conspicuity between VMIs and material images within a ±0. 5 margin, though the small sample limits generalization. Significance: This study demonstrates the feasibility of material decomposition using a physics-informed GAN model trained on realistic simulated data. The maps provided equivalent conspicuity under a clinically acceptable margin, with a significantly small number of images for interpretation.
Valand et al. (Thu,) studied this question.