Context
Breast cancer is the most represented cancer among women worldwide and a major cause of mortality. Early detection is key to the patient’s survival. This motivates vast coordinated screening of populations, which mostly uses projectional X-ray mammography. During the past years, research has extended its objectives beyond tumor detection and now put much efforts in risk assessment in order to intensify surveillance of women at higher risk of developing a breast cancer. It has been known for decades now that the parenchymal patterns visible in mammographies reflect cancer risk. Yet, identifying risky patterns in X-ray images is very difficult, even for experienced radiologists as the visual appearance of healthy dense tissues and disrupted tissues might be very similar, impairing their precise discrimination by eye.
Recent works showed that the microenvironment degradation, which impairs the suppression of emerging lesions and thus constitutes an important risk pattern, can be assessed by looking at the self-similarity, i.e., fractal, properties of mammographic textures. The major limitation of this approach is that self-similarity properties are highly sensitive to measurement noise in the X-ray mammograms, as quantified recently in the preliminary work. Precise observation of the fine properties of mammographic textures thus require to accurately denoise X-ray mammograms while preserving self-similarity.
Objective
The overarching goal of this PhD thesis is to develop non-Gaussian denoisers preserving self-similarity, i.e., fractal, properties of the textures present in the image, with the aim of analyzing X-ray mammograms through the prism of fractal analysis. The PhD candidate will investigate both model-based and data-driven approaches to image denoising. He/she will then combine optimization procedures and deep learning techniques to get accurate yet robust and interpretable procedures.
A detailed subject is available here: https://uncloud.univ-nantes.fr/index.php/s/ZqrKQ54RdHQF79E.
Organization and hosting institution
The PhD candidate will benefit from the financial support of the ANR-funded UNGRAM project to attend workshops and conferences, and from the rich scientific environment of both the UNGRAM consortium, gathering researchers from Nantes, Vannes and Tours as well as international collaborators in the United States and in Italy. He/she will also have the opportunity to collaborate with a postdoc as well as two interns to be recruited in the project.
The recruited candidate will be hired by the Centre National de la Recherche Scientifique (CNRS) in the framework of ANR grant UNGRAM holded by B. Pascal. He/she will integrate the Laboratoire des Sciences du Numérique de Nantes LS2N, in the Signal, Image and Sound (SIMS) team and work on the campus of Centrale Nantes, a top-level engineering school.
Applications
The recruited PhD student is expected to be at ease with the basic concepts of statistics, machine learning, linear algebra and convex optimization. Knowledge in convex nonsmooth optimization, with a special focus on proximity operators, would be a plus. Good programming skills in Python are required, in particular in the use of PyTorch environment, as well as good written and oral communication skills in English.
Candidates are requested to send a CV, master grades, references and motivations to barbara.pascal@cnrs.fr. The position is three year-long and can start anytime in 2027–administrative procedures need to be anticipated. Applications will be collected until a suitable candidate is found. Informal information requests are to be emailed to barbara.pascal@cnrs.fr.
