Soutenance de thèse de Pau Soler
Mardi 21 Mars à 11H00, Amphi Émeraude
Télécom Paris - 46, rue Barrault - 75013 Paris

Restauration Échographique Multi-vues appliquée à l’Imagerie 3D du Sein et du Coeur

Pau Soler
Mardi 21 Mars à 11H00
Télécom Paris, Barrault, Amphi Émeraude
Directeur de thèse
Membres du jury
  • Alain Herment (INSERM, Paris)
  • Alison Noble (University of Oxford)
  • Christian Barillot (INRIA/IRISA, Rennes)
  • Olivier Gérard (Philips Medical Systems, Suresnes)
  • Éric Saloux (CHU Caen)
3D ultrasound, spatial compounding, multichannel deconvolution, breast imaging, cardiac imaging.


Ultrasound echography is one of the most widely used medical imaging modalities because it is non-invasive, real-time and cost-effective. However, it suffers from some image quality defects, such as speckle noise, limited spatial resolution and angle dependent tissue contrast. Spatial compounding, which consists in averaging acquisitions from different angles, has shown good performances in reducing speckle in 2D imaging. In our thesis, we extend spatial compounding capabilities to also improve spatial resolution and angle dependency, particularly for 3D breast and cardiac imaging.

We propose new techniques to combine acquisitions from different angles, following two approaches: multiview deconvolution and multiview fusion. Multiview deconvolution consists in solving the inverse problem of estimating the original volume from the different acquisitions, by modeling the degradation as a convolution of the original tissue with a space-varying point spread function (PSF). We propose a technique to estimate the PSF based on subspace techniques, adding a priori knowledge of the geometrical relationship between the different acquisitions and shape constraints. The inverse problem is regularized with an edge-preserving functional adapted to speckle noise statistics, and solved iteratively. The second approach, multiview fusion, consists in detecting the features of interest in each acquisition to build a combined volume. We propose different fusion techniques in spatial, spectral and wavelet coefficient domains.

We applied the developed techniques to 3D ultrasound breast imaging. Volumes were obtained by scanning the tissue with a linear array attached to a robotic platform. A well-known problem in breast imaging in the limited elevational resolution of linear arrays. We overcame this problem by scanning the tissue at different angles, and combining those acquisitions. A non-rigid registration process was developed to guarantee the alignment of the different acquisitions. Multiview deconvolution methods showed best results with respect to improvement of spatial resolution and signal-to-noise ratio (SNR). This leads to the improvement of important parameters for clinical practice, such as tissue delineation and contrast resolution both on phantom and in vivo data.

Finally, we applied the developed techniques to real-time three-dimensional (RT3D) ultrasound cardiac imaging. This imaging modality can be further improved by increasing the field-of-view (FOV) and heart wall contrast. With this aim, different views were acquired through different acoustic windows. In order to combine such acquisitions, a robust rigid registration algorithm was developed. Multiview deconvolution showed slightly better results than other techniques, improving heart wall contrast and SNR. This leads to improvement of clinically relevant parameters, such as tissue delineation and heart wall contrast.

Page maintenue par le webmaster - 3 février 2010
© Télécom ParisTech/TSI 1998-2010