ESTRO 2024 - Abstract Book

S3911

Physics - Image acquisition and processing

ESTRO 2024

2. Tisell A, Lundberg P, Warntjes M, Testud F. 3D Quantitative MRI of the Brain: Effects of B1 Inhomogeneity in 3D-QALAS. ISMRM; 2021; Online.

3. Warntjes M, Engström M, Tisell A, Lundberg P. Modeling the Presence of Myelin and Edema in the Brain Based on Multi-Parametric Quantitative MRI. Frontiers in neurology. 2016;7:16.

4. Bledsoe JC. Effects of Cranial Radiation on Structural and Functional Brain Development in Pediatric Brain Tumors. Journal of Pediatric Neuropsychology. 2016;2(1):3-13.

5. Westin CF, Knutsson H, Pasternak O, Szczepankiewicz F, Özarslan E, van Westen D, et al. Q-space trajectory imaging for multidimensional diffusion MRI of the human brain. Neuroimage. 2016;135:345-62.

2326

Digital Poster

Brain MR-to-CT generation: comparison between supervised and unsupervised cGAN

Blanche Texier 1 , Adélie Queffélec 1 , Zahra Jenkal 1 , Cédric Hémon 1 , Yassin Kortli 1 , Safaa Tahri 1 , Songyue Han 1 , Renaud de Crevoisier 1,2 , Caroline Lafond 1,2 , Joël Castelli 1,2 , Anaïs Barateau 1,2 , Jean-Claude Nunes 1

1 Univ Rennes, LTSI INSERM UMR1099, Rennes, France. 2 Univ Rennes, Centre Eugène Marquis, Rennes, France

Purpose/Objective:

The current gold standard imaging for radiotherapy is computed tomography (CT). Nevertheless, MRI is being considered more and more as a suitable solution. Indeed, it offers better soft tissue contrast and the trend is emphasised by the development of MRI-Linacs. However, MRI suffers from the lack of electronic density information essential for an accurate dose calculation. As there is no direct link between MRI intensities and Hounsfield Units (HU), synthetic CT (sCT) generation by Deep Learning (DL) appears to be a solution because it combines MRI content information with CT intensities in HU [1]. DL architectures are now based on the Generative Adversarial Networks (GAN) [2] due to their accuracy, often used in a 2D supervised context. Indeed, supervised models are highly precise on a paired database and 2D networks have a lower computational cost than 3D methods. However, 2D methods generate reconstruction artefacts that can be avoided with 3D architectures. Finally, supervised DL approaches require a perfectly registered large amount of paired data, this can be avoided with unsupervised learning [1].

The main contribution of this work consists of a comparison between supervised and unsupervised conditional GAN approaches for sCT generation in the brain area.

Material/Methods:

In this study, 51 patients with brain tumours were included. For each patient, a CT scan (SIEMENS SOMATOM Confidence) in treatment position with a thermoformed mask and an MRI were acquired. The T1 MRIs were acquired on a 1.5T GE Optima MR450w (47 patients), a 3T Philips Ingenia (3 patients) and on a 1.5T SIEMENS MAGNETOM Amira (1 patient).