Accueil › Animation Scientifique › Armando GARCIA HERNANDEZ, PhD

Armando GARCIA HERNANDEZ will defend his Thesis defense entitled "Synthetic Computed Tomography Generation for Magnetic Resonance-only Radiotherapy" on Tuesday, September 13th at 11:00 a.m. in the Pierre Cotton Room of Institute Fresnel.

His thesis was supervised by Mouloud ADEL from Team GSM of Institute Fresnel and Pierre FAU from Institut Paoli-Calmettes.

 A Zoom link will be provided soon

Abstract : Radiotherapy (RT) is one of the most important cancer treatment techniques available. It is based on the delivery of high doses of ionizing radiation beams to targeted volumes in the body composed of cancer cells to damage their DNA. Recent developments in RT and medical imaging techniques, such as Magnetic Resonance Imaging (MRI), have resulted in the conception of hybrid systems that integrate MRI scanners with RT Linear accelerators (Linacs) into MR-Linacs. These systems have improved the performance of treatments allowing for novel methods such as Magnetic resonance image guided radiotherapy (MRgRT) and Adaptive radiotherapy (ART).
Computed Tomography (CT) is the gold standard in the field of RT as it provides the necessary electron density information required for dose calculation. Although MRI offers a better soft tissue contrast which helps in the identification and localization of tumors and organs at risk (OAR), CT is still needed. The use of both modalities, CT and MRI, demands an image registration stage which can introduce potential systematic errors in the dose calculation, as both images are acquired on different machines at different times.
A solution is the generation of synthetic CT (sCT) images from MR images. The generated sCT images provide the missing electron density data necessary for dose estimation that is lacking from the MRI. By replacing the CT images with the sCT counterparts, the whole CT acquisition process could be avoided, leading to a potential reduction in cost, resources and radiation for the patient. This thesis deals with the generation of sCT images for MR-only RT treatment planning.
Image quality is of great importance for RT treatment planning. MR-Linac systems such as the MRIdian from ViewRay use a low magnetic field of 0.35 Teslas (T) for its integrated MRI scanner to reduce the effect of the magnetic field on the delivered dose. However, as signal-to-noise ratio (SNR) is proportional to magnetic field strength, this system results in lower SNR images compared to other clinical MRI systems with strenghts of 1.5 T up to 3 T. Therefore, we proposed a denoising pre-processing step of low-field MR images using a Denoising AutoEncoder. Our proposed method outperformed other state-of-the-art techniques for MRI denoising.
Neural network architectures such as : U-Net and GAN have been studied for the generation of mainly brain and pelvic sCT images with seldom studies focusing on the abdomen, where strong physiological motion occurs. Additionally, most of these studies have been performed with 1.5 T MR images. We propose the generation of abdominal sCT images from 0.35 T MR images using both U-Net and GAN architectures. Bulk density images denoted as Simplified-sCT images have also been generated with both neural network architectures. These images were composed of only 6 densities with the objective of simplifying the image generation task to only 6 gray-levels as opposed to generating a whole range gray-scale image.
Generated sCT images were evaluated in terms of image quality metrics by calculating the Mean error (ME) and Mean absolute error (MAE) between them and the reference CT images. Dose accuracy was also compared between dosimetry plans obtained from the sCT images and the reference CT-based dose plans. Dose distributions were calculated on the generated images and compared with the reference CT-based dose plans. Dose volume histogram (DVH) and gamma passing rate (GPR) scores were compared to determine dose accuracy.
We have shown that sCT images generated from low-field MR images are capable of producing dosimetry plans with a 95% accuracy and dose differences of <1.65%. These images were generated on average in less than 2.5 seconds, making them suitable for use in ART where dose re-optimization
occurs while the patient remains on the table. Further tests are necessary to improve the performance of sCT image generation and ease their integration into the RT workflow.

Keywords : Radiotherapy, Synthetic CT, Deep Learning, MR-only treatment planning, Low-field MRI