Background: Magnetic resonance (MR)-only radiotherapy eliminates computed tomography (CT) images and replaces them with synthetic CT (sCT) images. In this work, we present sCT image generation from head-and-neck MR images and the corresponding geometric and dosimetric analysis of organ-at-risk (OAR) segmentations in CT and sCT images. Methods: Using a deep learning algorithm based on the diffusion process, we trained a model on 42 pairs of CT and MR images and applied it to 12 test pairs to generate corresponding sCT images conditioned on MR inputs from random noise. We then automatically segmented 12 different OARs in the CT and sCT images, and performed geometric (computing the spatial alignment of the obtained segmentations in images) and dosimetric (computing the received radiation doses of the obtained segmentations from the corresponding radiation dose distribution plans) evaluation. Results: The generated sCT images are comparable in quality to those from existing methods: mean absolute error of 43.4 HU, peak signal-to-noise ratio of 30.6 dB and structural similarity index of 91.2%. Geometric analysis of OAR segmentations showed a high agreement between CT and sCT images: Dice similarity coefficient of 87.9% and 95th percentile Hausdorff distance of 2.2 mm. Similarly, dosimetric analysis showed slight differences in radiation doses between CT and sCT images: a relative difference of 2.3% in the mean dose and 1.9% in the maximum dose. Conclusions: The generation of sCT images using the proposed methodology yields encouraging results, indicating that sCT images are viable alternatives to CT images from for geometric and dosimetric analysis of the corresponding OAR segmentations.
|
