Projet de thèse en Traitement du signal et des images
Sous la direction de Thomas Rodet et de Dominique Lesselier.
Thèses en préparation à Paris Saclay , dans le cadre de École doctorale Sciences et technologies de l'information et de la communication (Orsay, Essonne ; 2015-....) , en partenariat avec SATIE - Systèmes et Applications des Technologies de l'Information et de l'Energie (laboratoire) , MOSS - Méthodes et outils pour les Signaux et Systèmes (equipe de recherche) et de École normale supérieure Paris-Saclay (Cachan, Val-de-Marne) (établissement de préparation de la thèse) depuis le 01-09-2018 .
In the context of early detection of breast cancer and generally speaking of abnormalities by non-invasive/non-ionizing methods intended to work on an uncompressed breast, we wish to design/study algorithms of fusion/inversion of joint data. Electromagnetic and ultrasound modalities are easily deployed indeed in similar conditions (so-called free breast in coupling medium), and are complementary in terms of information, meaning local resolution and sensitivity to contrasts, and more generally about constituent media and distributions thereof. The objective is to model/characterize by simulation the potential performance of this bi-modal examination. Long-term aim is to propose a clinical system to detect early breast cancer. At this stage (proof-of-concept) we intend to use new methodologies of data fusion to design new multi-modality systems.
Early Breast Abnormalities Detection via Electromagnetic and Ultrasonic Joint Modalities
Start point. Detection of breast cancer so far is mostly via mammography, which suffers from compressed breasts (i.e., discomfort/pain), ionizing radiation deposited, not enabling to get the 3D structure, low contrast of tumors in mammograms, so small ones are not detected at early stage]. The initial idea was then to profit from strong electromagnetic contrasts of tumors, e.g., a microwave breast imaging system has been built without clinical exploitation attainable. Tridimensional microwave tomography involves illumination at different incidences & frequencies, and fields collected outside. From those, we can in principle retrieve within each 3-D voxel a complex-valued contrast, yet that suffers from reduced resolution if low frequencies, and if higher ones, from strong attenuations as waves penetrate farther and farther. Worst is that inversion here is non-linear, and even if using re-parameterization, we can make the problem multi-linear, matrices involved upon discretization are not hollow, which in particular limits the number of voxels that can be retrieved. Also, now in non-destructive testing framework, artificial objects to image often are involving a small number of materials (say, 2 to 4). Therein, tools address lack of high-frequency information using priors of images made of a small number of materials leading to constant piecewise ones. But breast structures in fine are too complex to use those, though works exist favoring constant images by optimization so that computing time fits clinical use]. Proposed approach. Its originality is in using two modalities: microwave and ultrasound. Simply said, the first yields low spatial frequencies of breasts and contrast in terms of tumors, the second yields higher frequencies and highly-resolved images. Now, to exploit at best the information from the two modalities, several conditions must be met as sketched below. First, measures in both --- ultrasound is already used --- must be almost simultaneous. Breasts are soft organs that deform in non-rigid manner between different examinations. Fusion between optical images and mammography where the breast is constrained has been proposed, but strong hypotheses are on internal structure deformation, and fusion merely determines the area within the unconstrained breast of what detected in mammography. Microwave and ultrasound modalities can cohabit in the same system without excessive interference. I.e., it appears enough to use a fluid enabling breast-sensor contact with good impedance properties (mechanical and electromagnetic). The second condition is of a parameterization enabling a relationship between the two modalities. To do so, we will consider that the images of the two modalities share same contours, this aspect being tackled via so-called line variables. Finally, an information fusion methodology must be used which is coherent with an inverse problem solution method, and a Bayesian methodology will well fit therein. Let us emphasize that we have already investigated already that type of fusion approach, in either X-ray or in fusion between x-ray (tomosynthesis) and microwave (typical radar approach). To avoid the limitation of resolution in microwave imaging caused by the relatively small number of illuminations that can be introduced because of the complexity of the direct model, we will use an adaptive grid of quadtree type, so that in particular homogeneous zones are modeled by a small number of voxels focusing onto the boundaries of objects (information that we will get from ultrasound data), this modeling being done thanks to the available expertise at the laboratory and partners. To recap. The project will lead to the development of a new multi-modality data fusion algorithm (microwave & ultrasound), which to best of our knowledge has not been conducted before. The performance of the approach will be evaluated on realistic numerical simulation data. The work is expected to render possible to get real gains in terms of resolution and contrast, noticing also that in the case of tumors of small sizes the bad resolution indeed reduces the contrast due to the partial volume effect (the volume of the tumor is embedded within the volume of the voxel, which is too large).