Image Projects DIGIT/IMA - January 2026
Master of Computer Science, Sorbonne Université
Table of Contents
- 1. Adapting VoxelMorph to compute cellular deformations in biological imaging
- 2. Image Segmentation of Marine Sediment in Electron Microscopy
- 3. AI-Based Digital Twins of Seafloor Environments
- 4. High-Fidelity 3D Meshing from Point Clouds for Marine Habitat Analysis
- 5. Development of a Virtual in game Assistant for Video Game Accessibility for Visually Impaired and Blind Users
- 6. Development of a Virtual Assistant for Video Game in game menu navigator Accessibility for Visually Impaired and Blind Users
- 7. Self-Supervised “Zero-Shot” SR vs Supervised SRResNet
- 8. Electron Back-Scattered Diffusion et indexation dictionnaire
- 9. Medical Image Segmentation and Classification Based on UNet Model
- 10. Traditional or generative method for spine image denoising?
- 11. Unsupervised segmentation on spinal and paraspinal regions
- 12. Evaluating Foundation Model Representations for Cell Segmentation and Classification in Microscopy Images
- 13. Optimisation de la résolution de cartographies chimiques pour les géosciences
1. Adapting VoxelMorph to compute cellular deformations in biological imaging
Supervisor: mailto:josephine.lahmani@pasteur.fr
Number of students: 2
The amount of deformations a cell undergoes during migration is an important biomarker of cancer metastasis. These deformations can be quantified by computing the strain field from videos of moving biological cells.
We propose to adapt the VoxelMorph algorithm for such purposes. Namely, we will train the VoxelMorph algorithm on provided videos of deforming glioblastoma cells. Unlike the original paper, the loss function would have to incorporate an additional term of the Hessian of the displacement field. A systematic comparison of this method with others previously developed will be performed.
Reference [1] Balakrishnan, G., Zhao, A., Sabuncu, M. R., Guttag, J., & Dalca, A. V. (2019). VoxelMorph: A learning framework for deformable medical image registration. **IEEE Transactions on Medical Imaging, 38**(8), 1788–1800.
2. Image Segmentation of Marine Sediment in Electron Microscopy
Supervisor: mailto:Fabrice.Minoletti@sorbonne-universite.fr
Number of students: 2
Description:
The composition of deep marine sediments responds to climatic variations and thus serves as an indicator of both past and current changes in Earth's climate. Quantifying the content of these sediments involves analyzing digital images acquired through optical microscopy. Thresholding and detecting regions of interest is a fundamental step prior to sample quantification and can significantly impact the obtained results.
Prerequisites: Image processing, thresholding methods, shape detection methods. The work will be conducted in Python.
Task Requirements:
The objective is to test and compare several thresholding methods (for RoI detection). Once the method(s) is (are) identified (bibliographic study), it (they) will be implemented in Python and tested on reference image libraries. Using the Python library openCV is recommended The development of a simple graphical user interface (GUI) to compare results will be appreciated.
Image example are available here: https://www.dropbox.com/scl/fo/nx3anoypb04dzhgxrbw03/AMWQr_KaGRgO83LoCjGcwBA?rlkey=xz2scfq1fihrr0drzapa53izn&dl=0
3. AI-Based Digital Twins of Seafloor Environments
Supervisors: mailto:loica.avanthey@epita.fr, mailto:laurent.beaudoin@epital.fr
Number of students: 2
See project details.
4. High-Fidelity 3D Meshing from Point Clouds for Marine Habitat Analysis
Supervisors: mailto:loica.avanthey@epita.fr, mailto:laurent.beaudoin@epital.fr
Number of students: 2
See project details.
5. Development of a Virtual in game Assistant for Video Game Accessibility for Visually Impaired and Blind Users
Supervisor: mailto:fabien.verite@sorbonne-universite.fr
Number of students: 2
See projet details.
6. Development of a Virtual Assistant for Video Game in game menu navigator Accessibility for Visually Impaired and Blind Users
Supervisor: mailto:fabien.verite@sorbonne-universite.fr
Number of students: 2
See projet details.
7. Self-Supervised “Zero-Shot” SR vs Supervised SRResNet
Supervisor: mailto:ghassen.marrakchi@lipn.univ-paris13.fr
Number of students : 2
See projet details.
8. Electron Back-Scattered Diffusion et indexation dictionnaire
Supervisor: mailto:loic.labrousse@sorbonne-universite.fr
Number of students : 2
See project details.
9. Medical Image Segmentation and Classification Based on UNet Model
Supervisor: mailto:yuxuan.xi@lip6.fr
Number of students: 2
See project details.
10. Traditional or generative method for spine image denoising?
Supervisor: mailto:clara.bremond@inserm.fr
Number of students: 2
See project details.
11. Unsupervised segmentation on spinal and paraspinal regions
Supervisor: mailto:clara.bremond@inserm.fr
Number of students: 2
See project details.
12. Evaluating Foundation Model Representations for Cell Segmentation and Classification in Microscopy Images
Supervisor: mailto:vvadori@pasteur.fr
Number of students: 2
See project details.
13. Optimisation de la résolution de cartographies chimiques pour les géosciences
Supervisor: mailto:nicolas.rividi@sorbonne-universite.fr (CAMPARIS), mailto:guillaume.bonnet@sorbonne-universite.fr (ISTeP), mailto:dominique.bereziat@lip6.fr
Number of students: 2
L’analyse chimique in-situ est un outil commun des géosciences, permettant de déterminer les conditions de cristallisation et de réaction de minéraux, mais aussi de mobilité des éléments au sein des minéraux. L’amélioration des techniques analytiques dans le domaine des géosciences a démontré que la résolution d’analyse était un des facteurs limitants pour l’interprétation de phénomènes géologiques rapides ou intermittents. Aujourd’hui, des techniques fiables (sonde ionique, microscope et microsondes électroniques), permettent d’effectuer des analyses chimiques à des échelles inférieures au micromètre. Néanmoins, ces analyses sont beaucoup plus longues, notamment lorsqu’il s’agit de réaliser des cartographies chimiques à très haute résolution spatiale (0,3 µm/pixel).
Les algorithmes de super-résolution pourraient permettre d’optimiser le temps d’acquisition de ces cartographies chimiques en effectuant des acquisitions à faible résolution (1-5 µm/pixel) et d’accéder ensuite à une image plus finement résolue par calcul. Un enjeu dans ce cas concerne la fiabilité de l’information chimique créée.
L’objectif du stage est donc d’entraîner un modèle de super-résolution à partir d’un jeu de cartographies chimiques d’une même bibliothèque d’échantillons à différentes résolutions. Ce modèle sera ensuite testé sur des cartographies chimiques à faible résolution obtenues sur d’autres minéraux issus de contextes similaires, et ses résultats comparés aux cartographies chimiques acquises en haute résolution.