Going beyond voxel-wise deconvolution in perfusion MRI: learning and leveraging spatio-temporal regularities with the stU-Net
Résumé
We present a deep learning method able to accurately predict quantitative perfusion maps from raw perfusion sequences, even in low-dose regimes of gadolinium-based contrast agent. The proposed stU-Net architecture is composed of a 2dimensional decoder and a 3-dimensional encoder able to exploit spatio-temporal regularities by learning convolutional filters that jointly act on the spatial and temporal domains. The method is evaluated on a public data set containing 49 patients with brain tumors, using an original approach to simulate low-dose perfusion sequences.
Origine : Fichiers produits par l'(les) auteur(s)