ED Mathématiques et Informatique
Implementation of Artificial Neural Networks for the simulation of complex physics problems: application to plasma physics.
by Morad BEN TAYEB (IMB - Institut de Mathématiques de Bordeaux)
The defense will take place at 11h00 - salle de conférence Institut de Mathématiques de Bordeaux UMR 5251 Université de Bordeaux 351, Cours de la Libération F-33405 TALENCE
in front of the jury composed of
- Jean-Luc FEUGEAS - Directeur de recherche - Université de Bordeaux - Directeur de these
- Emmanuelle ABISSET-CHAVANNE - Professeure - Arts et Métiers ParisTech - École Nationale Supérieure d'Arts et Métiers - Rapporteur
- Marina OLAZABAL-LOUMé - Directrice de recherche - CEA/CESTA - Rapporteur
- Didier RAFFESTIN - Ingénieur de recherche - DAM/CELIA/CELIA - CoDirecteur de these
- Gaël POETTE - Ingénieur de recherche - CEA/CESTA - Examinateur
- Corisande LAMY - Ingénieure de recherche - CEA/DIF - Examinateur
- Bruno DUBROCA - Directeur de recherche - CEA LCTS - Examinateur
This work integrates AI into ICF modeling along two axes: surrogate neural models that accelerate nonlocal electron heat transport while preserving multi-physics fidelity, and generative AI to optimize laser drive and target designs. Nonlocal transport is implemented in Azathoth via a multi-group SNB correction to Spitzer–Harm and validated on temperature step tests; a neural surrogate trained on Azathoth data delivers ~100× faster inference with sub-1% temperature error in transients. In CHIC 2D, a Fourier Neural Operator learns the Te → SNL mapping with strong cross-grid generalization and >30× acceleration at 512×512 for smooth-gradient regimes. A VAE-based pipeline structures the latent space of laser profiles and, with a gain regressor, identifies high-yield regions; validated CHIC cases reach gain ~113 at ~460 kJ (shock-augmented-like at far lower energy) and ~117 at ~348 kJ with smoothed direct-drive profiles (no shock), promising for high-rep-rate systems. For helical-coil targets guiding post-TNSA protons, databases connecting geometry to performance (E99, E95, Qsup) highlight low-pitch/large-radius designs and the feasibility of short coils; architectural advances (dual encoders/decoders, dual latents, scalar integration) improve reconstruction and robustness. The thesis concludes with paths to 3D neural operators, geometry-aware operators, enriched multi-physics datasets, and experiment-informed generative optimization
ED Sciences de la Vie et de la Santé
Astrocytic SERGLYCIN exhibits ambivalent properties in the control of astrogliosis processes.
by Margaux LAISNE (Biologie des maladies cardiovasculaires)
The defense will take place at 14h00 - Salle de séminaire Inserm U1034, 1 avenue Magellan, 33600 Pessac
in front of the jury composed of
- Isabelle BARDOU - Maîtresse de conférences - Université de Caen - Rapporteur
- Baptiste LACOSTE - Full professor - University of Ottawa - Rapporteur
- Aude PANATIER - Directrice de recherche - Université de Bordeaux - Examinateur
- Carole ESCARTIN - Directrice de recherche - Université Paris-Saclay - Examinateur
Introduction : After central nervous system (CNS) insult, astrocytes become reactive, undergoing morphological and molecular changes named "astrogliosis". This phenomenon is either beneficial, promoting the astrocyte borders around CNS lesions that limits the size of lesions, the spread of inflammatory cells and preserve healthy tissues, or deleterious, with the production of factors leading to blood-brain barrier (BBB) opening. To characterize astrogliosis signature, we performed a RNA-sequencing on “non-stimulated” versus “reactive” human astrocytes (hRA) and identified Serglycin (SRGN), an intracellular proteoglycan, as upregulated in hRA. We confirmed the overexpression of SRGN in astrocytes in vivo in neurovascular unit from Experimental autoimmune encephalomyelitis induced mice, on brain section from mice induced with focal cortical lesion and on cortical sections of active lesion from multiple sclerosis patients. Aim : Our aim is to elucidate the contribution of astrocytic SRGN in the control of astrogliosis in response to a CNS injury. Methods : We subjected inducible astrocyte-specific SRGN knockout mice (SrgnACKO) to intracortical stereotaxic injection of NaCl to induce a focal cortical injury. We evaluated glial border formation at 2-, 4- and 7-days post injection (dpi). To determine the role of SRGN on astrogliosis and associated signaling pathways in vitro, we used gain- and loss-of function approaches on hRA treated or not with IL-1β to induce astrogliosis or incubated human brain microvascular cells (hBMC) with hRA conditioned medium. Results : We demonstrated that SRGN overexpression by reactive astrocytes favors the formation of a protective astrocytes border around tissue lesion after focal cortical injury. SrgnACKO mice show less astrocytic reactivity, an increased spread of inflammation and significant neuronal loss at 2- and 4-dpi. SRGN contributes to astrocyte hyperplasia, elongation and barrier properties of border-forming reactive astrocytes both in vivo after focal cortical injury and in vitro in a model of IL-1β-induced astrogliosis. In vitro, we demonstrated that SRGN controls actin cytoskeleton remodeling through its interaction with the receptor CD44 and activation of the p42-44 MAPK and Src kinase signaling pathways. Conversely, we highlighted in vitro that SRGN overexpression promotes a pro-inflammatory chemokine-rich secretum in reactive astrocytes that impairs BBB integrity but were unable to identify this effect in vivo following focal cortical injury. Conclusion : Collectively, our data demonstrated that SRGN-CD44 axis induces some beneficial effects of astrogliosis in response to CNS injury, favoring the formation of astrocytes border around lesion tissue after focal cortical injury thus limiting the post-injury lesion spread. However, SRGN expression in reactive astrocytes is also associated with deleterious effects of astrogliosis by promoting a chemokine-rich secretum deleterious for BBB integrity in vitro. Thus, we believe that the SRGN-CD44 axis in reactive astrocytes exerts dichotomous roles in astrogliosis effects depending on the context and raises exciting questions about the role of this signaling pathway in controlling reactive astrocytes inflammatory profile and its impact on neuropathology with a strong inflammatory component such as ischemic stroke or multiple sclerosis.
ED Sociétés, Politique, Santé Publique
Joint models and deep learning neural networks for the longitudinal analysis of mammography images in screen-detected breast cancer risk prediction
by Manel RAKEZ (Bordeaux Population Health Research Center)
The defense will take place at h00 - Amphi Louis - ISPED 146 rue Léo Saignat, 33076 Bordeaux Cedex
in front of the jury composed of
- Virginie RONDEAU - Directrice de recherche - Equipe BIOSTAT - BPH U1219 - Directeur de these
- Brice AMADEO - Maître de conférences - Equipe EPICENE - BPH U1219 - CoDirecteur de these
- Roch GIORGI - Professeur des universités - praticien hospitalier - UMR 1252 SESSTIM - Rapporteur
- Agathe GUILLOUX - Professeure des universités - INRIA-INSERM équipe HeKA - Rapporteur
- Suzette DELALOGE - Praticienne hospitalière - Département de Médecine Oncologique - Gustave Roussy - Examinateur
- Olivier HUMBERT - Professeur des universités - praticien hospitalier - UMR E4320 TIRO-MATOs - Examinateur
- Rodolphe THIéBAUT - Professeur des universités - praticien hospitalier - Equipe SISTM - BPH - Examinateur
Breast cancer is the most common cancer among women and a leading cause of death worldwide. Early detection and precise risk prediction remain critical for effective prevention and targeted screening. Yet, existing models rarely exploit the longitudinal information available from repeated mammograms in population-based screening programs. This thesis aimed to develop and evaluate methods that leverage longitudinal mammographic imaging to enhance individual risk prediction. Two complementary strategies were explored. The first approach combines deep learning with joint modeling to link quantitative imaging biomarkers with cancer risk over time. A convolutional neural network, based on a modified U-Net architecture, was developed to segment breast and dense tissue and estimate validated measures of mammographic density at each screening visit. These quantitative measures, established risk factors for breast cancer, were incorporated into a joint modeling framework to quantify their association with cancer occurrence over time using several latent structures. This approach, developed as the DeepJoint algorithm, generates individualized dynamic risk predictions while preserving statistical interpretability. The second approach, LongiMam, is an end-to-end deep learning model that predicts risk probabilities directly from longitudinal mammogram sequences, with no need for predefined biomarker extraction. By integrating convolutional and recurrent neural networks, it captures spatial and temporal patterns across repeated screenings. This data-driven method aims to maximize predictive accuracy, though with potentially less interpretability compared to biomarker-based models. Both methods were applied to a large American screening dataset, addressing complementary aspects of risk prediction. Our findings show that longitudinal modeling improves breast cancer risk prediction and support the use of repeated mammograms in risk assessment. All proposed tools are open source, ensuring transparency and reproducibility for future research.