Generative AI & transfer learning between imaging modalities in cytogenetic biol. dosimetry (Th25-17 H/F

Informations générales

Entité de rattachement

L'Autorité de sûreté nucléaire et de radioprotection est une autorité administrative indépendante créée par la loi du 21 mai 2024 relative à l'organisation de la gouvernance de la sûreté nucléaire et de la radioprotection pour répondre au défi de la relance de la filière nucléaire.

Elle assure, au nom de l'État, le contrôle des activités nucléaires civiles en France et remplit des missions d'expertise, de recherche, de formation et d'information des publics. L'ASNR est composée de fonctionnaires, d'agents de droit public et de salariés de droit privé.  

Référence

2025-1215  

Description du poste

Intitulé du poste

Generative AI & transfer learning between imaging modalities in cytogenetic biol. dosimetry (Th25-17 H/F

Type de contrat

Doctorat

Statut

Cadre

Disponibilité du poste

06/10/2025

Localisation du poste

Fontenay-aux-Roses

Environnement / Organisation / Contexte

In situations involving nuclear or medical accidents, or malicious radiological acts, knowing the level to which people have been exposed to ionising radiation is crucial in determining the appropriate medical response. As a complement to physical dosimetry, cytogenetic biological dosimetry is based on imaging techniques that enable the detection of radiation-induced chromosomal damage in circulating lymphocytes. This damage, known as chromosomal aberrations, is considered to be one of the most reliable biomarkers for estimating radiation exposure.
In previous work, computer vision and artificial intelligence models were developed to automatically detect chromosomal aberrations for two cytogenetic imaging modalities: Giemsa imaging and Fish3 imaging. In particular, this work revealed the possibility of using a ‘trans-modal' approach to move from one imaging modality to another, enabling the development of a computer vision model for a third cytogenetic imaging modality: multi-Fish (M-Fish).

Mission

This PhD thesis is a collaborative project between the ASNR's laboratory of radiobiology of accidental exposures and the INRIA's SAIRPICO project team. It proposes to deploy generative AI and transfer learning tools around three topics: 1) The development of an AI model for image pre-selection in Glemsa/Fish3 modalities, 2) Converting the automated count of aberrations in Fish3 imaging into a radiation dose, taking into account confounding factors and associated uncertainties, and finally 3) the development of a new AI model for the automated count of chromosomal aberrations in M-Fish imaging.


TASK 1 - Development of artificial intelligence models for the pre-selection of metaphase images in the Giernsa and Fish-3 modalities (from Q4 2025 to Q2 2026)
Classification convolutional neural networks will be trained on two databases (for each Giemsa and Fish modality) containing images classified into two labels by experts: non-exploitable images (around one thousand for each modality) or exploitable images (around five thousand). The similarity between the Giemsa images and the blue channel (DAPi) of the Fish3 images will be exploited using transfer learning methods in order to double the size of each training database.

TASK 2 - Dosimetric validation of an lA model for counting chromosomal aberrations in Fish3 imaging (from Q3 2026 to Q1 2027)
The aim is to combine two deep learning models, one for aberration detection and the other for calibration curve regression, in order to provide a personalised dose estimate and the associated uncertainties.
More specifically, the first model will be trained on a database augmented by image-to-image generative artificial intelligence in order to classify chromosomes and identify chromosomal aberrations in Fish3. The second model will model the baseline rate (in the absence of exposure to ionising radiation) of chromosomal aberrations by integrating individual demographic variables such as age, sex and smoking status.

TASK 3 - Development of an lA model for counting chromosome aberrations using M-Fish imaging (from Q2 2027 to Q2 2028)
Using 5 fluorescent probes (instead of three for the Fish3 technique), it is possible to stain each chromosome in a cell with a single colour using the M-Fish technique. This would allow all aberrations to be counted exhaustively. However, the analysis time is very long and impossible to deploy in an operational accident context. It also means that the number of training images available is fairly small (a few hundred).
This thesis project proposes a two-step approach to deal with this problem: use transfer learning techniques to develop generative models of chromosomal aberrations under M-Fish starting from models trained to generate aberrations under the fish3 modality, and then use these synthetic images as a data augmentation to develop an aberration classifier under M-Fish.

Profil recherché

Applicants should have a master degree in applied mathematics, data science or biostatistics. Previous experience in biological imaging would be an asset.The selected candidate will work in close interdisciplinary collaboration with scientists having expertise in applied mathematics and radiobiology. The position includes leading data analysis and manuscript writing in collaboration with the research team. The selected candidate will be encouraged to present the findings of the project at scientific conferences as well as to administrative authorities.

Télétravail

Occasionnel

Diversité

Diversity is one of the components of IASNR's CSR, HR and Quality of Working Life policies. We give equal consideration to all applications, without discrimination, to include all talents.

Whatever the differences, we aim to attract, integrate and retain our candidates and employees within an inclusive working environment.

ASNR pursues an active policy in favor of equal opportunities in the workplace and the employment of disabled people. If you have a disability, please let us know if you have any specific needs that we can take into account.

Localisation du poste

Localisation du poste

Europe, France, Ile-de-France, Hauts-de-Seine (92)

Critères candidat

Langues

• Anglais (2- Niveau professionnel)
• Français (2- Niveau professionnel)