MR image synthesis using Riemannian geometry constrained in VAE

Jannane Nada (1) , Sébastien Bougleux (2, 3) , Jérôme Lapuyade-Lahorgue (1) , Su Ruan (1) , Fethi Ghazouani (1)

1 QuantIF-LITIS - Equipe Quantification en Imagerie Fonctionnelle
2 Equipe Image - Laboratoire GREYC - UMR6072
3 UNICAEN - Université de Caen Normandie

Deep generative models are used to learn nonlinear data distributions across a set of latent variables and a generator function which maps the latent variables to the input space. In this paper, we propose a deep generative model based on variational auto-encoders (VAE) with Riemannian geometry for synthetizing MRI images. Riemannian geometry is used as a constraint to extract the most relevant and structuring features in latent space. This approach is applied on Brats20 data. Quantitative comparisons between a classical VAE and our method are carried out to evaluate the quality of the generated images. The results demonstrate the superiority of our approach.

Mots clés: Deep generative models; Riemannian geometry; VAE; Data augmentation; Latent space

Domaines: Sciences du Vivant [q-bio]

Jannane Nada, Sébastien Bougleux, Jérôme Lapuyade-Lahorgue, Su Ruan, Fethi Ghazouani. MR image synthesis using Riemannian geometry constrained in VAE. 16th IEEE International Conference on Signal Processing (ICSP), IEEE Beijing Section; Beijing Jiaotong University, Oct 2022, Beijing, China. pp.485-488, ⟨10.1109/ICSP56322.2022.9965357⟩. ⟨hal-03842257⟩