Stanford University Stanford University, CA, United States
B. Li1, S. Ye1, L. Zhu2, Y. Chen1, L. Liu3, L. Yu2, and L. Xing1; 1Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA, 2Department of Statistics and Actuarial Science, The University of Hong Kong, Hong Kong, China, 3Department of Radiation Oncology, Stanford University, Stanford, CA
Purpose/Objective(s): Dual-energy CT (DECT) provides a valuable tool for a variety of clinical applications, such as accurate proton therapy dose calculation, better tissue characterization, reduced imaging artifact and material quantification. However, the high cost associated with the system hinders its widespread clinical implementation in radiation therapy (RT) treatment planning and onboard CBCT imaging. We propose a multi-sinogram learning (MSL) framework that extract and leverage the shared information between energies for DECT reconstruction using single-energy sinograms, making DECT accessible to clinics with single-energy CT or CBCT systems. Materials/
Methods: We used the dual-energy CT dataset from the AAPM spectral CT grand challenge, where training and testing dataset consisted of 400 and 100 images. We generated the dual energy sinograms by forward projecting the DECT images. The MSL involves a transformer block for intra- and inter-energy features learning. A physics-informed encoder-decoder architecture was used for domain transfer between sinogram and image. We reconstructed DECT images from single energy sinograms from the test dataset with MSL and performed image domain two material decomposition. The results were compared quantitatively with that from FBP reconstruction of dual-energy sinograms using mean-squared error (MSE) and peak signal-to-noise ratio (PSNR). Results: On the test dataset, the MSL-generated DECT images from high-energy sinogram with lower average MSE of 466.4 and higher PSNR of 45.49 compared to 722.5 and 41.03 from FBP reconstructions of dual-energy sinograms. Accurate material-specific images were obtained from material decomposition of MSL reconstructed DECT images with lower average MSE of 0.0017 and higher PSNR of 55.71 compared to 0.0283 and 30.94 from material decomposition of the FBP reconstructed DECT images. Similar results were achieved when only using low-energy sinograms with MSL. Conclusion: MSL accurately reconstructed DECT images from single-energy sinograms, thereby enabling broadly used conventional single-energy CT systems to incorporate DECT functionalities.
