Mots-Clés

AI Morphogenesis deep-learning embryo biology imaging

Description

Context

Artificial intelligence (AI), and in particular deep learning, is transforming image analysis, especially in the biomedical field. However, current methods still struggle to efficiently process large and complex datasets such as 3D+time microscopy recordings of biological processes, which have dramatically improved in recent years.
Next-generation microscopes can now capture in real time the full 3D dynamics of whole embryos. These datasets have sufficient resolution to follow embryonic development cell by cell with unprecedented precision. Yet their richness also poses a major challenge: imaging a single embryo, containing thousands of dynamically interacting cells, can generate several terabytes of data.
Deep learning is becoming fundamental to analyzing these datasets, particularly for:
• Automatically segmenting cells, i.e. distinguishing and delineating each individual cell within the volume.
• Tracking cell trajectories over time, including daughter cells after division.
• Objectively comparing embryonic development across individuals, in order to understand developmental dynamics and the structure of inter-individual variability, at multiple scales (global, tissue, cellular).
This PhD project is positioned precisely at the frontier between artificial intelligence and developmental biology. Advanced knowledge in developmental biology is not required.

Objectives

The PhD candidate will develop new AI methods to:
1. Segmentation and cell tracking in 3D+time: Design end-to-end architectures integrating both segmentation and tracking, in order to achieve reconstructions that accurately capture individual variability.
2. Unsupervised morphological representation: From imaging data and spatio-temporal reconstructions of embryonic morphology, develop a compact latent space that enables objective comparison of embryos, independent of their size or orientation, and detection of typical or atypical developmental trajectories.
3. Integration into a unified framework: Produce a comprehensive methodology for the comparative and quantitative analysis of complex embryonic data.
4. Dissemination of results: Deliver an open-source software tool accessible to the international scientific community.

Methodology

The thesis will explore two complementary axes:
• Axis 1: Spatio-temporal segmentation and tracking: Development of sparse convolutional neural networks (sparse CNNs) combined with attention mechanisms and hybrid CNN–Transformer architectures. This approach will explicitly model cellular dynamics (divisions, collective movements) and reinforce the temporal coherence of reconstructions.

• Axis 2: Unsupervised learning and latent representations: Use of 3D autoencoders and implicit representations (Neural Radiance Fields, DeepSDF) to model embryonic forms. The goal is to filter out irrelevant variations (noise, orientation, artifacts) while preserving the essential morphological structures.
The models will be trained and validated on a unique dataset of ascidian embryos already acquired. Ascidians, a sister group of vertebrates, were chosen as a model organism because their embryogenesis is rapid and highly stereotyped.

Expected results

• A robust and objective methodology for comparing complex embryonic morphologies and their dynamics.
• An innovative neural network dedicated to segmentation and tracking in 3D+time.
• An open-source software tool for embryo analysis, made available to the community.
• International publications in AI, biomedical imaging, and developmental biology.
This project will represent a paradigm shift in the automated, quantitative study of embryogenesis, with potential applications to other biological systems.

Candidate profile

• Master’s degree (MSc) or engineering diploma in computer science, applied mathematics, bioinformatics, biomedical imaging, or equivalent.
• Skills in deep learning (PyTorch, TensorFlow), programming (Python required), and image analysis.
• Strong interest in interdisciplinary research and good command of scientific English.
This project is aimed at students passionate about applying advanced AI methods to an académic research project and eager to contribute to a better understanding of the fundamental mechanisms of animal embryonic development.