Pour vous authentifier, privilégiez eduGAIN / To authenticate, prefer eduGAIN
10 September 2021
Europe/Paris timezone

Program of the day (all presentations are available at the bottom of this page)

  • 10h00 - 10h15 : Welcome
  • 10h15 - 10h30 : D. Ullmo : presentation of Institut Pascal
  • 10h30 - 11h15 : L. Jacques, UCLouvain : The importance of phase in complex compressive sensing. In this talk, we consider the estimation of a sparse (or low-complexity) signal from the phase of complex random measurements, a "phase-only compressive sensing" (PO-CS) scenario. This study is motivated by the fixed range of such measurements, which can for instance ease further data quantization methods, with potential applications in radar, computed tomography or computational imaging. With high probability and up to a global unknown amplitude, we show that perfect signal recovery is possible if the sensing matrix is a complex Gaussian random matrix and the number of measurements is large compared to the signal sparsity. Our approach consists in recasting the (non-linear) PO-CS scheme as a linear compressive sensing model. We built it from a signal normalization constraint and a phase-consistency constraint. Practically, we achieve stable and robust signal direction estimation from the basis pursuit denoising program. Numerically, robust signal direction estimation is reached at about twice the number of measurements needed for signal recovery in compressive sensing (joint work with T. Feuillen) 
  • 11h30 - 12h00 : M. de Bruijne, University Medical Center Rotterdam: Weak labels and adversaries in medical imaging. Machine learning approaches, and especially deep neural networks, have had tremendous success in medical imaging in the past few years. Automated, quantitative image analysis with convolutional neural networks is now in many cases as accurate as the assessment of an expert observer. A major factor hampering the adoption of these techniques in practice is that it can be very expensive, time-consuming, or even impossible to obtain sufficiently many representative and well-annotated training images. On the other hand, weaker, image-level labels are often readily available in the form of a radiologist’s assessment of the presence or absence of certain abnormalities. In this talk, we will discuss various approaches to make use of such data. If time permits, I will also touch upon our recent work on adversarial attacks in medical image analysis. In specific, we have studied the effects of model pretraining and of variations in training data on the vulnerability of medical image classification networks to black-box adversarial attacks that manipulate the input image with the aim to fool the network into making wrong decisions.
  • 12h00 - 12h30 : L. Oudre, ENS Paris-Saclay : Change-point detection with application to human gait analysis. Non-stationary time series can be found in numerious fields such as industrial monitoring or healthcare. When dealing with such signals, change-point detection appears as an important preprocessing task, which consists in finding the temporal boundaries of homogeneous sub-signals. This talk will present an overview of  offline change-point detection methods in an unified framework, along with recent supervised approaches to calibrate them. The different algorithms will be applied on physiological data, by using the ruptures package.
     

Lunch break

 

  • 14h00 - 14h45 : M. Unser, EPFL : Deep splines. Our intent is to demonstrate the optimality of splines for the resolution of inverse problems in imaging and the design of deep neural networks. To that end, we first present a representer theorem that states that the extremal points of a broad class of linear inverse problems with a generalized total-variation constraint are adaptive splines whose type is linked to the underlying regularization operator. When the latter is the 2nd-order derivative operator, then the optimal reconstruction is an adaptive piecewise-linear spline with the smallest possible number of breakpoints. These splines are intrinsically sparse, and hence compatible with the kind of formulation (and algorithms) used in compressed sensing. We then show that our results are relevant to the investigation of neural networks as well. In particular, they yield a functional interpretation of shallow, infinite-width ReLU neural nets. Sparse adaptive splines also turn out to be ideally suited for the specification of deep neural networks with free-form activations.
  • 15h00 - 15h30 : Z. Ramzi, CEA Saclay & INRIA : Practical use of Jean Zay supercomputer. Deep Learning (DL) has become the main research avenue in many computational fields, and MRI follows this trend. However, training and evaluating DL models requires huge amounts of expensive computing resources. In an effort to mutualize these resources for researchers in France, GENCI has developed Jean Zay, a supercomputer installed at IDRIS (CNRS, Orsay). In this presentation, we will discuss some broad aspects of Jean Zay for computational scientists and illustrate them with a personal journey.
  • 15h30 - 16h00 : J. Schnabel, King’s College London: AI-enabled medical imaging. Artificial intelligence, in particular from the class of machine / deep learning, has shown great promise for application in medical imaging. However, the success of AI-based techniques is limited by the availability and quality of the training data. A common approach is to train methods on well annotated and curated databases of high-quality image acquisitions, which then may fail on real patient cases in a hospital setting. Another problematic is the lack of sufficient numbers of clinical label annotations in the training data, or example for early markers of disease. In this talk I will present some of our recent approaches that aim to address some of these challenges, by using AI as an enabling technique for improved image reconstruction, realistic data augmentation and further downstream tasks.
  • 16h00 - 16h30 : S. Farrens, CEA Saclay : Python Sparse data Analysis Package. During this talk I will present the PySAP (Python Sparse data Analysis Package) package, a multidisciplinary image processing tool developed in collaboration between astrophysicists and biomedical imaging experts at CEA Paris-Saclay through the COSMIC project. I will provide some background on the core mathematical tools implemented in PySAP as well as demonstrating some of the diverse applications such as galaxy image deconvolution and magnetic resonance image reconstruction. I will also endeavour to relate how we managed the development of this package. Finally, I will conclude by sharing our plans for the future of PySAP and how this may impact other imaging domains.
  • 16h30 : Conclusion