Science

I am working, generally speaking, on statistical image processing. Applications are (for now) related to astronomical and medical images. Insights of my previous and current research activities are given below.

Analysis of cloud satellite images (current)

During my post-doc, I work on time series of multi-spectral images in order to analyze cloud aggregation. These works concern the fields of particle filtering and sparse optimization.


My PhD thesis dealt with the statistical analysis of hyperspectral astronomical images, aiming at revealing the presence of faint, distant (several billion light-year) objects in deep field images. These images were acquired with the MUSE instrument, which is set in the ESO VLT facility in Paranal, Chile. This PhD thesis stemmed several research topics, detailed below.

Synthetic view of all segmentations made by the convolutional pairwise Markov field model and the spatial triplet Markov tree model (see below). The background is the MUSE image averaged over bandwidth. On the foreground, the detected haloes are color-coded depending on their distance (redder=farther, up to 13 billions light-years).

Convolutional Pairwise Markov Fields (ca. 2016-2017)

In these works, we formulate the detection problem as an instance of Bayesian segmentation within a Pairwise Markov field model, accounting for the convolution in images. This approach was applied on the detection of faint extended sources in very noisy astronomical hyperspectral images (see image above).

A journal paper is in preparation on this topic.

Spatial Triplet Markov Trees (ca. 2017)

In these works, we investigate the use of the auxiliary process from the Triplet Markov modeling to allow the use of both hierarchical relations between the tree resolutions and within-resolutions relations to allow the segmentation of large-scale structures in extremely noisy images. This approach proved to be more efficient that Markov-field based method for the detection in very noisy  images (about -20dB).

For now this work is only published in the GRETSI conference (preprint, in French), and a journal paper is in preparation.

 Oriented Triplet Markov Field (ca. 2016)

This work is part of the effort to detect wide and oriented features in MUSE deep field observations, which would be the so-called cosmic web. The initial purpose of Oriented Triplet Markov Field  (or OTMF) was to provide a tool to detect such features. While there is yet no cosmic web detection proof with OTMF in MUSE images, we observed that OTMF are efficient in segmenting images presenting directional features.
This work was presented in IEEE WHISPERS 2016 (preprint),  and a journal publication has been submitted.

Segmentations instance of an image presenting linear features. Left: original image. Middle left: segmentation using a classical Hidden Markov Field model. Middle right and right: joint segmentation of image classes and local orientations.

Statistical detection of faint Lyman-alpha haloes (ca. 2015)

This work deals with the analysis of MUSE hyperspectral images to detect Lyman-alpha haloes, which are large and faint gas clouds orbiting around distant galaxies. The method we propose is based on refinement of GLR testing, accounting for specificities of the target signal. More specifically, we design a GLR accounting for the spectral shape, spatial field spread function, and consistency between spectra. The output of this work may provide for the first time a statistical mapping of these features.
These works were presented at the ICASSP 2016 conference (doi:10.1109/ICASSP.2016.7472005), and a complete journal publication is being processed. A demonstration code will soon be available !

Halo detection example. Red contour/average spectrum are the brightest detection (mostly related to the galaxy) and blue contour/average spectrum represent the extended faint detection. The background image is the hyperspectral image averaged over wavelength. The gray spectra is an instance of noisy MUSE spectra.

Collaborations

As a member of the MUSICOS team and the European MUSE consortium, I had the chance to contribute to some other works :

  • Extended Lyman alpha haloes around individual high-redshift galaxies revealed by MUSE, in Astronomy and Astrophysics (doi:10.1051/0004-6361/201527384).
  • R. Bacon et al., The MUSE Hubble Ultra Deep Field Survey. I. Survey description, data reduction, and source detection., accepted in july 2017 in Astronomy and Astrophysics, doi: https://doi.org/10.1051/0004-6361/201730833

 


Vertebrae segmentation in CT images (2014)

Segmentations instances on a full spine.

This work was done during my master’s internship between February and July, 2014. The purpose of this work was to provide an efficient segmentation method to address vertebrae in CT scan volumes, originating from daily practice cases. This implies that the vertebra of interest may not be healthy and “anatomically standard”, because patient which actually need vertebra scan are mostly not young nor healthy.
In this context, we developed a robust method aiming at segmenting vertebrae, regardless of their healthiness, actual shape or position in the spine.

These work are reported in the journal paper Vertebra Segmentation Based on a 2-step Refinement (doi:10.1186/s40244-016-0018-0) ; and presented at the IPTA international conference as well as in the GRETSI French conference.
A demo code will also be available.