Feasibility of Using Deep Learning to Generate Dual-Energy CT from 120-kV CT

Purpose Deep learning (DL) has been applied to generate a high-kV (e.g., 140 kV) computed tomography (CT) image from its low-kV (e.g., 80 or 100 kV) CT image. This indicates that dual-energy CT (DECT) analysis can be performed without using a DECT scanner. However, CT images are typically acquired at 120 kV instead of 80 and 100 kV. In this study, we investigate whether DL has the ability to generate both 80- and 140-kV CT images from 120-kV CT images. Methods We recruited ninety-eight patients who underwent brain DECT scans (80 kV/Sn140 kV). We emulated 120-kV CT images by a linear blend of 30% 80-kV and 70% 140-kV CT images. Thus, an additional 120-kV acquisition was not required. We trained a U-Net convolutional neural network to generate both 80- and 140-kV CT images from 120-kV CT images. Results We observed that the DL-based DECT images were visually similar to the true (original) DECT images. Moreover, the difference in mean CT number between the true and DL-based DECT images was less than 1 HU for brain, fat, muscle, and cerebrospinal fluid. There were no statistically significant differences in attenuation between the true and DL-based DECT images ( p > 0.05) for the four studied tissues. Conclusion Our preliminary results demonstrate that DL has the potential to generate both 80- and 140-kV CT images using 120-kV CT images.