Abstract in Englisch:

In this paper, we propose 2D MRI-based pseudo-CT (pCT) generation approaches that are inspired by U-Net and generative adversarial networks (GANs) and that additionally utilize coarse bone segmentation guided attention (SGA) mechanisms for better image synthesis. We first introduce and formulate SGA and its extended version (E-SGA), then we embed them into our baseline U-Net and conditional Wasserstein GAN (cWGAN) architectures. Since manual bone annotations are expensive, we derive coarse bone segmentations from CT/pCT images via thresholding and utilize them during the training phase to guide image-to-image translation attention networks. For inference, no additional segmentations are required. The performance of the proposed methods regarding the image generation quality is evaluated on the publicly available RIRE data set. Since MR and CT image pairs in this data set are not correctly aligned with each other, we also briefly describe the applied image registration procedure. The results of our experiments are compared to baseline U-Net and conditional Wasserstein GAN implementations and demonstrate improvements for bone regions.