Deep learning methods for virtual monoenergetic imaging from spectral CT - Archive ouverte HAL Accéder directement au contenu
Pré-Publication, Document De Travail Année : 2023

Deep learning methods for virtual monoenergetic imaging from spectral CT

Suzanne Bussod
Nicolas Ducros
Philippe Douek
Christine Chappard

Résumé

Spectral photon counting computed tomography (CT) is a X-ray imaging modality that acquires energy-resolved data thanks to photon counting detectors that sort photons depending on their energy. This allows to decompose the object into its material constituents or to reconstruct virtual monoenergetic images. In this paper, we address for the first time the reconstruction of virtual monoenergetic images from spectral CT measurements, which is a non linear inverse problem, focusing on the application to knee osteoarthrisis. While traditional methods are based on the inversion of a physical model, deep learning methods have recently demonstrated their ability to solve inverse problems. In this work, we propose several physics-informed deep learning strategies for virtual monoenergetic image reconstruction. We consider four different reconstruction algorithms for the recovery of virtual monoenergetic images in the projection and in the image domain. All of our algorithms include a variant of the U-net convolutional neural network. The proposed algorithms were trained and evaluated on the spectral CT data simulated from realistic knee phantoms generated from synchrotron radiation CT. They were also compared to a Gauss-Newton algorithm that minimized a cost function with a hand-crafted regularization term. Finally, our algorithms were applied to an experimental knee data set acquired on a clinical spectral CT scanner. We found that the proposed approaches provide virtual monoenergetic images with improved mean squared errors and structural similarities, compared to the Gauss-Newton method. Moreover, the image-domain network improved the mean squared error by a factor of two, compared to the projection-domain network. In both simulated and experimental data of osteoarthritis knees, we found that the cartilage was visible with naked eye on the virtual monoenergetic images reconstructed by our methods. The proposed deep learning networks outperformed the Gauss-Newton algorithm in the projection domain. Among deep reconstruction strategies, we found that the image-domain direct virtual monoenergetic reconstruction performs the best. They also allow for the direct visualization of the cartilage, which is essential for the assessment of cartilage integrity.
Fichier principal
Vignette du fichier
preprint_HAL.pdf (2.29 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04346653 , version 1 (15-12-2023)

Licence

Paternité

Identifiants

  • HAL Id : hal-04346653 , version 1

Citer

Suzanne Bussod, Nicolas Ducros, Juan F P J Abascal, Philippe Douek, Christine Chappard, et al.. Deep learning methods for virtual monoenergetic imaging from spectral CT. 2023. ⟨hal-04346653⟩
72 Consultations
55 Téléchargements

Partager

Gmail Facebook X LinkedIn More