Back to overview

Regularized Linear Inverse Problems in Diffusion Magnetic Resonance and Ultrasound Imaging

English title Regularized Linear Inverse Problems in Diffusion Magnetic Resonance and Ultrasound Imaging
Applicant Thiran Jean-Philippe
Number 175974
Funding scheme Project funding (Div. I-III)
Research institution Laboratoire de traitement des signaux 5 EPFL - STI - IEL - LTS5
Institution of higher education EPF Lausanne - EPFL
Main discipline Information Technology
Start/End 01.11.2017 - 31.10.2021
Approved amount 750'000.00
Show all

Keywords (6)

ultrasound imaging; inverse problems; diffusion MRI; medical imaging; brain connectivity analysis; deep networks

Lay Summary (French)

L’imagerie médicale occupe une place de choix dans le diagnostic médical, et cette place est appelée à croitre avec l’évolution annoncée de la médecine personnalisée. Dans ce contexte, l’acquisition, la reconstruction et l’analyse des images sont des composantes essentielles de cette évolution.Dans ce projet, nous proposons de développer le cadre général des problèmes inverses linéaires régularisés pour la reconstruction des images médicales, et nous en développerons les composants principaux, dans le contexte de deux modalités d’imagerie majeures : l’imagerie IRM de diffusion et l’imagerie ultrason.
Lay summary

IRM de diffusion : depuis près de 15 ans, le LTS5 de l’EPFL se positionne à l’avant-garde de l’analyse de la connectivité cérébrale par IRM de diffusion. Récemment, nous avons proposé une série de contributions méthodologiques importantes permettant l’estimation robuste de la microstructure de la substance blanche. Dans ce projet, nous continuerons cet effort afin d’obtenir des estimations optimales et validées.

Imagerie ultrason : récemment, le LTS5 a proposé de considérer la reconstruction d’images échographiques comme un problème inverse linéaire. Dans ce projet, nous allons continuer et étendre ces travaux à l’imagerie 3D et nous allons également aborder la question de la complexité computationnelle de ces reconstructions au travers d’approches très originale reposant sur des réseaux de neurones profonds.

A la fin de ce projet, nous aurons donc développé les méthodologies essentielles pour reconstruire efficacement les informations issues de ces deux modalités majeures d’imagerie médicale. Par là, nous contribuerons à l’amélioration des dispositifs d’imagerie médicale, répondant ainsi aux besoins de la communauté de recherche biomédicale.

Direct link to Lay Summary Last update: 29.09.2017

Responsible applicant and co-applicants


Project partner

Associated projects

Number Title Start Funding scheme
157063 Towards micro-structure-based tractography for quantitative brain connectivity analysis 01.10.2014 Project funding (Div. I-III)
138311 Advanced signal processing on the sphere for high angular resolution diffusion magnetic resonance imaging 01.04.2012 Project funding (Div. I-III)
170758 High-End 3D Ultrasound Open Research Platform 01.12.2016 R'EQUIP
170873 Exploring brain communication pathways by combining diffusion based quantitative structural connectivity and EEG source imaging : application to physiological and epileptic networks 01.03.2017 Sinergia


Medical imaging occupies a place of choice in medical diagnosis, and this place will keep growing in the future the development of new imaging modalities and with the increasing need for personalized medicine. In this context, medical image acquisition, reconstruction and analysis are key technical components, subject to intense research all over the World, to provide the medical community with the most advanced and robust methods to extract information from the fantastic existing and future image modalities.In this project, we will promote the framework of regularized linear inverse problems in medical image acquisition and reconstruction, and develop some of its essential components, in the context of two very attractive medical imaging modalities: diffusion Magnetic Resonance Imaging (dMRI) and Ultrasound (US) imaging. This project is built as a continuation and extension of two important research lines pursued at the Signal Processing Laboratory (LTS5) of EPFL, and addresses key questions in these domains in a unified methodological framework. Indeed, in our previous works both in brain connectivity analysis by dMRI and in US image reconstruction, we proposed to formulate the data/image reconstruction aspects as regularized linear inverse problems and obtained significant preliminary results.Diffusion MRI: For the last 15 years, LTS5 has pioneered the field of brain connectivity analysis by dMRI, establishing the principle of MR connectomics. Recently, we developed an additional major contribution to the field, by reformulating the dMRI white matter microstructure estimation problems into linear inverse problems, for both microstructure imaging and microstructure informed tractography. In this project, we will continue this effort by investigating some of the key issues to obtain optimal estimation, namely dictionary design and learning as well as validation. Ultrasound imaging: Although now a mature field, medical US remains a modality supported by intense research and with extensive diagnostic and therapeutic indications in routine clinical use worldwide. Even if the research is very active, the basic component of US imaging, i.e. the beamforming method called Delay-and-Sum (DAS), has been largely untouched for several decades. While being very effective thanks to its simplicity, this method is largely suboptimal. Recently, LTS5 has developed the idea of addressing US image reconstruction as a linear inverse problem. Preliminary results already demonstrate outstanding performances in 2D imaging, both in image quality and data reduction. In this project, we will continue this effort by addressing some of the key aspects of this new paradigm, that will be required for this innovative method to become effective and have the expected impact. We will first extend it to 3D US imaging, where our framework has the potential to have the strongest impact, by enabling a high image quality while drastically reducing the data requirement and therefore making this technology appropriate for a much larger diffusion in the medical community. Secondly, we will address the computational complexity of US image reconstruction through regularized inverse problems by exploring and developing the remarkably promising idea of exploiting deep neural networks for image reconstruction.At the end of this project, we will thus have developed a series of new medical image acquisition and reconstruction methods, that will allow an optimal exploitation of these two amazing technologies: dMRI and US imaging. Finally, by improving medical imaging technologies, we will play our role of engineers, and contribute to the development of enhanced medical imaging devices, serving the need of the biomedical community, to ultimately provide new tools and methods for better understanding the human body and improved treatments for patients.