Truth-based Physics Informed Deep Learning Model for Material Composition Estimation in Spectral CT.

Abstract

Background: Spectral CT data from Photon-Counting CT (PCCT) can be used for material decomposition through various methods, including model-based approaches like Maximum Likelihood Estimation (MLE) and deep learning models. However, a significant challenge is the lack of accurate ground truth data, leading to a reliance on Hounsfield Unit (HU) values to estimate material compositions that are sensitive to noise and detector non-idealities. Moreover, most deep learning models employ pixel-wise loss functions that fail to incorporate underlying physics principles, limiting the accuracy of material quantification.Purpose: To develop and validate a physics-informed deep learning model, trained on ground truth data, to perform material decomposition of spectral CT images into density (ρ) and effective atomic number (Zeff) maps.Materials and Methods: The training dataset consisted of simulated abdominal PCCT scans generated from 32 computational phantoms and their corresponding ground truth. The images were obtained at two different clinical dose levels (CTDIvol of 1.5 and 6 mGy) at four sets of low and high detector energy thresholds, including (20,50) keV, (20,60) keV, (20,70) keV and (20,80) keV with four different iodinated contrast agent concentrations and three reconstruction kernels. The image dataset was used to train a Generative Adversarial Network (GAN) without and with a physics-informed regularization loss. The model performance was evaluated on a test dataset comprising 16 computational phantoms and validated on 10 clinical cases with different contrast and reconstruction conditions. The model output was also compared to ρ and Zeff values obtained using an analytical method.Results: With physics-informed regularization, the trained model showed superior performance. NRMSE decreased by 1.31% and 3.1%, PSNR increased by 2.1dB and 3.75dB, and SSIM was greater than 0.99 for the ρ and Zeff maps, respectively. Our model also showed reliable performance on the clinical data with a maximum RMSE of 3.87% for the predicted value of ρ for blood without arterial contrast and a maximum variation of 0.2% and 3.7% in values of ρ and Zeff, respectively, for images reconstructed with different reconstruction kernels. The model also showed superior performance in comparison to ρ and Zeff obtained numerically. NRMSE decreased by 13.81% and 33.72%, PSNR increased by 20.35dB and 19.78dB, and SSIM was greater than 0.98 for the ρ and Zeff maps, respectively.Conclusion: This study demonstrates the feasibility of material decomposition using a physics-informed GAN model trained on data from a virtual imaging trials platform, addressing limited access to training datasets with ground truth. The results also indicate reliable rendering of the ρ and Zeff maps of the clinical cases and a superior performance of the model compared to the analytical method.