Image synthesis with deep convolutional generative adversarial networks for material decomposition in dual-energy CT from a kilovoltage CT

Abstract

Generative Adversarial Networks (GANs) have been widely used and it is expected to use for the clinical examination and image. The objective of the current study was to synthesize material decomposition images of bone-water (bone(water)) and fat-water (fat(water)) reconstructed from dual-energy computed tomography (DECT) using an equivalent kilovoltage-CT (kV-CT) image and a deep conditional GAN. The effective atomic number images were reconstructed using DECT. We used 18,084 images of 28 patients divided into two datasets: the training data for the model included 16,146 images (20 patients) and the test data for evaluation included 1938 images (8 patients). Image prediction frameworks of the equivalent single energy CT images at 120 kVp to the effective atomic number images were created. The image-synthesis framework was based on a CNN with a generator and discriminator. The mean absolute error (MAE), relative mean square error (MSE), relative root mean square error (RMSE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and mutual information (MI) were evaluated. The Hounsfield unit (HU) difference between the synthesized and reference material decomposition images of bone(water) and fat(water) were within 5.3 HU and 20.3 HU, respectively. The average MAE, MSE, RMSE, SSIM, and MI of the synthesized and reference material decomposition of the bone(water) images were 0.8, 1.3, 0.9, 0.9, 55.3, and 0.8, respectively. The average MAE, MSE, RMSE, SSIM, and MI of the synthesized and reference material decomposition of the fat(water) images were 0.0, 0.0, 0.1, 0.9, 72.1, and 1.4, respectively. The proposed model can act as a suitable alternative to the existing methods for the reconstruction of material decomposition images of bone(water) and fat(water) reconstructed via DECT from kV-CT.