Image processing in biology and health

signal / image processing tools for biomedical applications.

On the image processing side, an efficient reconstruction procedure mixing on non-smooth convex optimization to balance data fidelity and regularization provides an efficient reconstruction procedure for high resolution image from low resolution records.

On the signal processing side, statistical tools are customized to analyze scale invariance in heart rate variability (notably in the context of intrapartum fetal heart beat analysis) as well as in infraslow brain activity. It notably permits to relate scale invariance in background brain activity to functional connectivity and to quantify task-rest modulation of brain scaling.

ML and SIP for studies in Health studies.

We also develop structured domain processing (structured as graphs, and even in graph-time-frequency spaces) to process brain activity networks. Espcially, these are used in ANR project SEIZURE to process MEG and iEEEG data and extract, from time-frequency features and epileptogenic network, the seizure onset zone.  

When the Covid19 pandemic broke out, we develop novel estimates for the space-time dynamics of the pandemics, quantified by the reproduction number. We elaborated on works in proximal optimization of  piecewise-smooth denoising, and this provided robust estimates despite the limited quality of the available new infection daily counts. Now, in ANR OptiMoCSI, optimization and Monte Carlo sampling are leveraged for a better processing of epidemics data, with quantification of uncertainties. We shared to a general, non expert, audience via interactive maps. 

We are leading the statistical and ML processing for the project VORTEX (of PEPR MIE), about the analysis of the Volatilome to develop a non-invasive diagnostic method of respiratory infections, based on mass spectrometry of patients' exhaled air; it relies on structured domain analyses of the signatures of organic compounds specific to a pathogen or to the immune response of patients. For that, our efforts are also devoted in the ShapeMed@Lyon project VOCALISE to the analysis with ML and statistics of the data ofthe COVIDAIR study.

Microscopy image analysis

Structured illumination microscopy increases fluorescence microscopy resolution without constraint on protein marking. This modality is based on the acqui- sition on several low resolution modulated images fol- lowed by a post-processing that aims at reconstruct- ing the high resolution image. The first contribution concerns the estimation of the modulation parameters from the low resolution modulated images while the second contribution aims at providing an efficient reconstruction procedure based on non-smooth convex optimization. Such a framework allows us to deal with a variational approach where the data fidelity term and the regularization term are fitted to the degradation model (Poisson noise) and to the data (filaments that models actin and microtubules, spots such as single molecules or vesicles).