Universities and Institutes of France

Universities and Institutes of France

Postdoc In Computer Science, Image Data Processing


Closing in 14 days

12 Sep 2023

Job Information


Université de Strasbourg


Direction de la Recherche

Research Field

Computer science

Researcher Profile

First Stage Researcher (R1)

Recognised Researcher (R2)



Application Deadline

12 Oct 2023 - 23:59 (Europe/Paris)

Type of Contract


Job Status


Hours Per Week


Offer Starting Date

1 Nov 2023

Is the job funded through the EU Research Framework Programme?

Not funded by an EU programme

Is the Job related to staff position within a Research Infrastructure?


Offer Description

1. Position identification

Title of post : postdoc in computer science, image data processing

Type of contract : fiwed term contract

Category (A,B or C) :

Contract/project period : 12 mouths Expected date of employment : November 2023

Proportion of work : 100%

Workplace : ICUBE

Desired level of education : PhD

Experience required : 1 to 3 years

Contact(s) for information on the position (identity, position, e-mail address, telephone) : Hassen Drira, professor, [email protected], 0 3 68 85 44 90

Date of publication : September 12, 2023

Closing date for the receipt of applications : October 12, 2023

2. Research project or operation

The objective of this project is to model 3D shape sequences from medical imaging (shape of a mouse brain, human hippocampus, etc.) in the shape spaces of parametrized surfaces, which will allow an invariant representation to the position in space, the scale and reparameterization. This representation will allow to calculate optimal deformations, distances and statistics between shapes, independent of rotation, translation, scale and re- parameterization. Each element of the sequence resulting from a frame will be considered as an element of the space of the forms of the surfaces and the 3D sequences resulting from medical data (MRI for example) will be considered as trajectories on the nonlinear space, called shape manifold. This representation will allow to develop a mathematical framework allowing to calculate an invariant distance to the execution rate, optimal deformations and statistics between sequences of shapes. It will then be necessary to extend deep learning architectures to these non-linear spaces. A difficult aspect is the extension of the convolution operator on manifolds in order to design deep architectures based on CNN (Convolutional Neural Networks). Further study will also be considered in order to establish the properties of the learning procedure on the variety (stability, convergence, impact of the metric, etc.).

3. Activities

  • Description of the research activities :
  • Shape is an important feature of objects and can be immensely useful in characterizing objects for the purpose of detection, tracking, classification, and recognition. As an example, it plays an important role in medical image analysis where advances in non-invasive imaging technology have enabled researchers to study biological variations of anatomical structures. Studying shapes of 3D anatomical structures in the brain is of particular interest because many diseases can potentially be linked to alterations of these shapes. There has been a significant amount of research and activity in the general area of shape analysis. By shape analysis we mean a set of tools for comparing, matching, deforming, and modelling shapes. The main differences amongst different tools proposed so far lie in the mathematical representations and metrics used in the analysis. For example, in shape analysis of planar objects (objects in 2D images), a variety of mathematical representations, including binary images, sampled points (active shape models 1), ordered points (lamdmark-based shape analysis 2), medial axes 3, level sets 4, and others, have been used. These different representations, along with their corresponding choices of metrics, lead to different solutions with their respective strengths and limitations. Traditional approaches often quantify data images issued from humans such data in vector space. Even deep learning models have impressively surpassed conventional models while assuming an underlying Euclidean structure of the space of the data, this assumption may not always be valid for data extracted from images. Actually, the topology of a given space characterizes the proximity between data and plays a vital role in pattern recognition. Pattern analysis takes place in the context of data lying in some inherent geometrical structure. Simply ignoring the geometrical aspect, or naively treating the space as Euclidean, may cause undesired effects.

    Fortunately, there have been increasing efforts applied to pattern analysis on manifolds. To account for the geometry of the images, manifold learning techniques like ISOmetric Mapping (ISOMAP) 5 and Local Linear Embedding (LLE) 6 were introduced. These approaches are based on learning a mapping from ambiant space to intrinsic space so that nearby points remain near each other after a projection. Another school of thought is to represent images in an underlying parametrized space 7,8. This gives rise to the representation of matrix manifolds. While ISOMAP and LLE model a manifold through training data, the use of matrix manifolds derives from the properties of differential geometry. In particular, data may be viewed as elements in some parameter space in which the idiosyncratic aspects of the geometry of the data can be characterized using algebraic operations. In fact, image data are often seen as the orbit of elements under the action of matrix manifolds, e.g. rotation group. Matrix manifolds may be the natural representation for some computer vision applications. Pennec et al. 9,10,11 propose interesting modeling of medical data on manifolds by designing manifolds of the groups that act on the shapes, thus the geodesics and means calculated on the manifolds are not directly shapes.

    This effort gave rise to a set of new mathematical representations of human data and their animations over time, termed trajectories, which live in non- linear manifolds. Tacking the example of 12, the authors propose to represent skeletal motion as trajectories in the Special Euclidean (Lie) group SE(3)n. Tacking a different direction, 13 have extended the Kendall's shape theory to trajectories. SVM classifier is performed to assess the classification once the data are represented on the manifold. Nevertheless, only few efforts have been spent on applying deep learning on manifolds 14, 15. To the best to our knowledge, only one recent work 16 has proposed an End-to-end deep learning architecture on a finite dimensional shape space representing the human landmarks and no such architecture is proposed for continuous data representing dense data (as surfaces).

    Scientific objective and innovation GDeepM project aims at investigating the intersection of deep learning and Riemannian geometry for shapes issued from medical imaging by designing a manifold representing 3D dynamic dense data and design deep learning architecture on it. The major innovation of this proposal is to model the 3D non rigid shapes issued from medical imaging (brain, hippocampe, heart, etc) in Riemannian manifolds and apply deep learning on these non-Euclidean spaces. A challenging aspect is the extension of the convolution operator on Riemannian manifolds in order to design deep CNN-based architectures. A more in-depth study will be also considered in order to establish the properties of the learning procedure on the Riemannian manifold (stability, convergence, impact of the Riemannian metric,...). Each shape (one frame) will be considered as an element of a non linear space (manifold) and the sequence can be considered as a trajectory on this manifold. Grounding on this precise space-time representation, one can first develop a mathematical framework with various tools as (rate invariant) metrics, geodesics, derivatives and some statistics (means) on the underlying spaces. In collaboration with Pr Frederic Blanc (Neurologist, Geriatrics in IMIS team in ICube) we propose to analyze the shape of hippocampus as use case, analyze the shape of the hippocampal surfaces of people with Alzheimer's disease and healthy ones. The deep learning based learning will be adapted to Riemannian manifold to train the quantity of abnormality in a given sequence of shapes, predict the evolution of shape change, classify shapes onto classes of pathology labelled by doctors, etc. To the best of our knowledge, only few papers have already proposed an adaptation of convolution to riemannian geometry 17. We note that despite the expertise on Riemannian geometry applied to computer vision developed by the Project Investigator, the present topic is novel enough from methodological perspectives. Actually, the intersection of riemannian geometry of shape space and deep learning is a new axis of research.

    The project is organized in 3 main workpackages :

  • Design manifold of and represent the 3D dynamic data as trajectories on this manifold, this include the definition of intrinsic calculations (geodesics, Frechet mean, rate invariant comparison of trajectories, etc).
  • Design Deep learning layers on the manifold: this workpackage include the designing of novel convolution operations adapted on the on linear space and study the convergence and stability of such architecture and the impact of the riemannien metric.
  • Built end-to-end deep learning architecture based on tools proposed in the previous workpackage inorder to classify, predict the sequences of 3D shapes
  • 1 Laga, H., Padilla, M., Jermyn, I. H., Kurtek, S., Bennamoun, M., & Srivastava, A. (2022). 4d atlas: Statistical analysis of the spatio-temporal variability in longitudinal 3d shape data. IEEE Transactions on Pattern Analysis and Machine Intelligence.

    2 Jermyn, I. H., Kurtek, S., Laga, H., & Srivastava, A. (2017). Elastic shape analysis of three-dimensional objects. Synthesis Lectures on Computer Vision , 12 (1), 1-185.

    3 Joshi, S. H., Xie, Q., Kurtek, S., Srivastava, A., & Laga, H. (2016). Surface shape Morphometry for hippocampal modeling in Alzheimer's disease. In 2016 International Conference on Digital Image Computing: Techniques and Applications (DICTA) (pp. 1-8). IEEE.

    4 Xie, Q., Jermyn, I., Kurtek, S., & Srivastava, A. (2014, September). Numerical inversion of srnfs for efficient elastic shape analysis of star- shaped objects. In European conference on computer vision (pp. 485-499). Springer, Cham.

    5 Pottier, C., Hannequin, D., Coutant, S. et al. High frequency of potentially pathogenic SORL1 mutations in autosomal dominant early-onset Alzheimer disease. Mol Psychiatry 17, 875–879 (2012). https: // doi.org/10.1038/mp.2012.15

    6 Sayette et al., Contribution to Alzheimer's disease risk of rare variants in TREM2, SORL1, and ABCA7 in 1779 cases and 1273 controls, Neurobiology of Aging, Volume 59, 2017, Pages 220.e1-220.e9, ISSN 0197-4580.

    7 Philippi N., Botzung A., Noblet V., Rousseau F., Després O., Cretin B., Kremer S., Blanc F., Manning L., Impaired emotional autobiographical memory associated with right amygdalar-hippocampal atrophy in Alzheimer's disease patients,Frontiers in Aging Neuroscience, Volume 7, 2015 , ISSN=1663-4365

    8 Schueller E., Paiva I., Blanc F., Wang X., Cassel J., Boutillier A., Bousiges O.,,

    Dysregulation of histone acetylation pathways in hippocampus and frontal cortex of Alzheimer's disease patients, European Neuropsychopharmacology, Volume 33, 2020,

    Pages 101-116,ISSN 0924-977X

    4. Skills

  • Qualifications/knowledge : PhD in computer science, image processing, knowledge in deep learning, experience in medical imaging processing
  • Operational skills/expertise : Programming in python, Pytorch, tensorflow
  • Personal qualities : Work in team, share knowledge
  • 5. Environment and context of work

  • Presentation of the laboratory/unity :
  • The laboratory brings together two scientific communities in equal parts at the interface between the digital world and the physical world, thus giving it a unique configuration. With nearly 650 members, it is a major research force on the Strasbourg site. Federated by imagery, ICube's preferred fields of application are engineering for health, the environment and sustainable development.

  • Hierarchical relationship : Hassen DRIRA
  • Requirements

    Research Field

    Computer science

    Education Level

    PhD or equivalent

    Internal Application form(s) needed

    Drira RecrutementPostDoctorantsAnnexe1G.pdf


    (106.51 KB - PDF)


    Additional Information Work Location(s)

    Number of offers available



    Laboratoire des sciences de l'ingénieur, de l'informatique et de l'imagerie – ICUBE, UMR 7357






    Where to apply


    [email protected]





    https:// www. unistra.fr


    4 rue Blaise Pascal

    Postal Code



    Job details


    Postdoc In Computer Science, Image Data Processing


    Universities and Institutes of France




    September 13, 2023

    Application deadline

    October 12, 2023

    Job type



    Computer Science,Engineering