This study focuses on the implementation of diffusion models for the synthesis of computed tomography (CT) images from magnetic resonance (MR) images in the head and neck region.
Radiotherapy for head and neck cancers requires precise localization of tumors and critical organs (also known as organs-at-risk, OARs) based on imaging for the planning of radiation treatment. Traditionally, CT and MR images are used for this purpose. Especially for the delineation of bone tissues and other structures, CT images offer good contrast while also containing electron density information, which is crucial for the creation of radiation dose distribution maps. While CT images allow accurate measurement of radiation attenuation, MR images offer better contrast of soft tissues, which aids in improved localization of tumors and OARs.
Within the scope of this master's thesis, we explore the possibility of using MR-only radiotherapy planning as an alternative to the traditional CT-based approach. One advantage of this approach is the elimination of exposure to ionizing radiation, an inherent aspect of CT imaging. This reduction is particularly significant for children, given the increased risk of developing cancer later in life. We focus on the development and evaluation of artificial intelligence models, particularly diffusion neural networks, for the synthesis of synthetic CT images from MR images.
To evaluate the quality of the generated synthetic images, we employed various metrics, including structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), Fréchet inception distance (FID) and mean absolute error (MAE). Qualitative analysis demonstrated that the synthetic CT images resemble real CT images, a finding supported by quantitative results, where we achieved the following average values: 92.2 % SSIM, 33.1 dB PSNR, 3.6 FID, and 35.3 HU MAE. Additionally, we calculated quantitative metrics on electron density images and compared the performance of OAR segmentation between synthetic and real CT images.
The main contributions of this work are: the development and implementation of diffusion models for CT image synthesis from MR images, specifically for the head and neck region; a detailed analysis of the impact of different model settings and three-dimensional image assembly techniques on the quality of generated images; an empirical evaluation of the usability of synthetic CT images in the segmentation of critical anatomical structures, confirming their potential for clinical application.
We conclude that MR-only radiotherapy planning, supported by advanced generative artificial intelligence models, represents a promising alternative to the traditional CT-based approach.
|