ED Mathématiques et Informatique
Exploring the internal organization of tumor tissue through volumetric imaging, numerical methods, and deep learning
by Florian ROBERT (IMB - Institut de Mathématiques de Bordeaux)
The defense will take place at 14h00 - Salle des Actes 351 cours de la Libération, Bâtiment A33, 33400 Talence
in front of the jury composed of
- Baudouin DENIS DE SENNEVILLE - Directeur de recherche - Université de Bordeaux - Directeur de these
- Karin PERNET-GALLAY - Ingénieure de recherche - Université de Grenoble - Rapporteur
- Charles KERVRANN - Directeur de recherche - Centre INRIA de l'Université de Rennes - Rapporteur
- Nataliya SOKOLOVSKA - Professeure des universités - Sorbonne Université - Examinateur
- Christophe GROSSET - Directeur de recherche - Université de Bordeaux - CoDirecteur de these
- Christèle ETCHEGARAY - Chargée de recherche - Université de Bordeaux - Examinateur
One of the major challenges in pediatric oncology is to design more effective treatments, particularly for metastatic, recurrent, or resistant forms, while reducing side effects. To achieve this, a better understanding of the biological mechanisms and the functioning of tumor tissues is essential. Among the avenues currently being explored, the internal morphological organization of tumor tissues could play an important role in how cells respond to treatments and, consequently, in their effectiveness. It is in this context that this thesis is situated, focusing on the study of the bioarchitectural organization of hepatoblastoma tissues, the most common malignant liver tumor in children. Recent advances in three-dimensional imaging offer new opportunities to characterize in detail the cellular and subcellular structures within biological tissues. In particular, serial block-face scanning electron microscopy makes it possible to acquire anisotropic volumetric images at high resolution, well suited to the study of large tissue samples. The objective of this work is to transform these complex data into an accurate digital representation of the tumor tissue, through the construction of a digital twin. This construction is based on two complementary steps. The first part of my thesis consists in developing automated segmentation tools to identify and individualize the biological structures of interest within the volume, notably cells, nuclei, nucleoli, mitochondrial networks, lipid droplets, blood capillaries, and hemorrhagic areas. The second part aims to exploit this segmentation to analyze the internal organization of the tissue through the extraction of bioarchitectural parameters. My thesis work made it possible to introduce 35 parameters grouped into five main categories: size, shape, distances between structures, orientation, and texture. The analysis of these high-dimensional data relies on unsupervised machine learning methods, making it possible to identify complex and non-trivial patterns of organization within tumor tissues. All of this work is part of a digital modeling approach to tumor architecture at the cellular and subcellular scales. This approach makes it possible to explore the individual behaviors and interactions between structures, with the aim of formulating, confirming, or refuting the biological hypotheses underlying these tumor phenomena. It constitutes a complete methodological framework for the automated, quantitative, and interpretable exploration of relevant biological tissues. By offering unprecedented access to the internal organization of tissues, it has the potential to contribute to a better understanding of tumors, to improve diagnosis, to refine therapeutic protocols, and, ultimately, to optimize the care of children with cancer.
Artificial Intelligence-based non-invasive Neurotechnologies for adaptive and personalized post-stroke motor rehabilitation
by David TROCELLIER (LaBRI - Laboratoire Bordelais de Recherche en Informatique)
The defense will take place at 9h00 - Ada lovelace Inria Talence, 200 avenue de la vielle tour,
in front of the jury composed of
- Fabien LOTTE - Directeur de recherche - Centre Inria de l'Université de Bordeaux - Directeur de these
- Bernard N'KAOUA - Professeur des universités - Université de Bordeaux -Bordeaux population health - CoDirecteur de these
- Fabrizio DE VICO FALLANI - Directeur de recherche - Inria, Institut du cerveau - Examinateur
- Laurent BOUGRAIN - Maître de conférences - Université de Lorraine / LORIA (Laboratoire Lorrain de Recherche en Informatique et ses Applications) - Examinateur
- Guan CUNTAI - Full professor - Nanyang Technological University - Rapporteur
- Tetiana AKSENOVA - Directrice de recherche - Leti CEA - Rapporteur
Brain-Computer Interfaces (BCI) enable users to send commands to a computer using only their brain activity. One of the most studied paradigms in BCI is motor imagery (MI), where users imagine specific limb movements without performing them. This is possible because imagining a movement activates neuronal patterns similar to those involved in actual movement execution. MI signals can be decoded using machine learning (ML) or deep learning (DL) algorithms, which are trained to extract the neural features corresponding to these imagined movements. However, current machine learning techniques often lack robustness. A significant variability in BCI performance is observed across users, with 30% of individuals unable to achieve reliable control, as well as within-user variability. This performance variability can arise from subject-specific factors, such as individual traits and the ability to perform the mental task, as well as session-specific factors, such as the quality of the recorded electroencephalographic (EEG) signal and fluctuations in users' mental states (e.g., fatigue or motivation). In this thesis, we aimed to enhance the robustness of BCI classification algorithms by addressing these sources of variability. We propose 3 contributions to respond to this question. We first validated several neurophysiological factors of BCI performance variability using a large open-access dataset. We validate the impact of those factors in order to be able in a second step to create ML/DL models invariant to them. We then conducted a systematic review of classification algorithms that explicitly take into account these variability factors, either to improve classification accuracy or to achieve invariance to these factors. We proposed a modified version of the Linear Discriminant Analysis (LDA) classifier designed to be invariant to specific neurophysiological factors correlated with BCI performance or subjective mental states. Lastly, we evaluate the impact of DL algorithms in a cross-subject and within-subject training approach. We observe that training the model from a large range of subjects allows to increase BCI performance as well as being more robust to cross-subject variability. In the second part of this thesis, we study MI-BCI variability in a more applicative context, for motor rehabilitation for stroke patients. Indeed, performing MI exercises can enhance neuroplasticity, and BCIs can provide continuous sensory feedback and promoting motor recovery. Despite this potential, current BCI-based rehabilitation therapies remain limited, largely due to algorithmic robustness issues as stroke also induces cross and within subject variability. To improve the effectiveness of BCI-based motor rehabilitation, we first conducted a synthesis of the various factors that may predict the success of rehabilitation. These included predictors of BCI control performance, and predictors of motor recovery after stroke, both with and without BCI intervention. Based on this analysis, we proposed an experimental protocol, to be conducted in collaboration with a hospital, aimed at validating these predictors in a clinical setting. Finally, we included a methodological reflection on best practices in EEG signal analysis. We demonstrated that visual cues can introduce biases in ML models, particularly DL architectures, when not properly controlled. Then, we evaluated how the choice of a baseline in event-related desynchronisation (ERD) analysis can significantly affect the interpretation of results. Through this work, we aim to better understand the factors related to MI-BCI performance variability and to develop ML methods that explicitly account for this variability. We also contributed to the experimental design of clinical trials to validate predictors of BCI-based motor rehabilitation. Lastly, we point out the importance of rigorous ML approaches that focus not only on classification accuracy but also on the neurophysiological features used.
ED Sciences Chimiques
Synthesis of dithienonaphthyridine, a sulfur and nitrogen analog of pyrene, and study of its derivatives as low band gap materials
by Debi Gordone BILAMBI (Institut des Sciences Moléculaires)
The defense will take place at 9h00 - salle de conférence Bâtiment A12 Bâtiment A12, Université de Bordeaux, 351 Cr de la Libération, 33405 Talence
in front of the jury composed of
- Yohann André Georges NICOLAS - Maître de conférences - Bordeaux INP, ISM, UMR 5255 - Directeur de these
- Thierry TOUPANCE - Professeur - Université de Bordeaux, ISM, UMR 5255 - CoDirecteur de these
- Laurence VIGNAU - Professeure - Bordeaux INP, IMS, UMR 5218 - Examinateur
- Nicolas LECLERC - Directeur de recherche - CNRS, ICPEES, UMR 7515 - Examinateur
- Sylvain ACHELLE - Professeur - Institut des Sciences Chimiques de Rennes, UMR 6226 - Rapporteur
- Christine LARTIGAU-DAGRON - Maîtresse de conférences - Université de Pau et des Pays de l'Adour (UPPA), IPREM UMR 5254 - Rapporteur
Silicon is an essential element in today's electronics. For example, the integrated circuits found on the boards of the various devices that surround us contain it. However, over the past two decades, organic semiconductor compounds compete with silicon and other inorganic materials in specific fields, such as displays, thanks to AMOLED technology. In commercial solar cells, silicon is also present, but this leads to certain disadvantages that organic materials could easily overcome, such as their energy-intensive process, their low-speed production process and their high density. In the field of solar cells, prospective and experimental studies have shown that polymers with low band gaps (LBG) are needed to obtain better photovoltaic efficiencies. A first chapter outlines the fundamental principles of organic electronics and the interest of LBG materials. Using examples, the tools for designing π-conjugated systems are explained, leading to the presentation of a new aromatic structure - dithienonaphthyridine (DTN) - made up of fused rings of donor thiophenes and acceptor pyridines. The remainder of the thesis details, for the first time, the synthesis, photophysical and electronic properties of DTN and its derivatives. In the second chapter, the study explores the methods for preparing the key DTN-NH synthon, and then its reactivity through usual reactions in the chemistry of π-conjugated systems. Halogenated, triflated and alkylated derivatives were obtained and optimizations of reaction conditions discussed. Improved synthesis protocols, regioselectivity of reactions and choice of alkyl chains have led to monomer units that are sufficiently pure and soluble to enable the creation of π-conjugated polymers for organic electronics. In the third chapter, the synthesis of oligomers and polymers of O-alkylated DTN derivatives are presented, exploring various methods such as nickel(0) reductive coupling, palladium-catalyzed C-H bond activation, Suzuki coupling and electrochemical polymerization, highlighting the challenges associated with monomer solubility and reactivity. DTN dimers were successfully obtained, as well as their brominated derivatives. These methods also led to the formation of soluble oligomer mixtures (from 2 to 7 units). The monomers were also coupled electrochemically to produce an insoluble material from the DTN derivative bearing n-hexyl chains In the final chapter, the compounds were characterized and studied via UV-vis absorption spectroscopy, cyclic voltammetry, TD-DFT calculations and single-crystal X-ray diffraction. The redox potentials obtained by cyclic voltammetry and the absorption bands determined by UV-vis agree with the results of DFT calculations. In addition, a discussion of the results obtained for oligomers and polymers shows that the extrapolation method, based on data from only 3 oligomers, is an effective predictor. Finally, the DTN homopolymer leads to the production of an LBG material. Thus, these results provide a solid basis for designing new π-conjugated materials from DTN
ED Sociétés, Politique, Santé Publique
Maternal occupational multi-exposures during pregnancy and foetal growth
by Marie TARTAGLIA (Bordeaux Population Health Research Center)
The defense will take place at 15h00 - Amphithéâtre Louis Université de Bordeaux, Campus Carreire, Bâtiment ISPED, 146 rue Léo Saignat, 33000 Bordeaux
in front of the jury composed of
- Fleur DELVA - Praticienne hospitalière - Université de Bordeaux - Directeur de these
- Yolande ESQUIROL - Maîtresse de conférences - praticienne hospitalière - Université de Toulouse - Rapporteur
- Anne CHANTRY - Professeure des universités - Université Paris Cité - Rapporteur
- Ronan GARLANTEZEC - Professeur des universités - praticien hospitalier - Université de Rennes - CoDirecteur de these
- Pascal GUENEL - Directeur de recherche émérite - Université Paris Saclay - Examinateur
- Jean François GEHANNO - Professeur des universités - praticien hospitalier - Université de Rouen - Examinateur
Many pregnant women are simultaneously exposed to multiple occupational factors that may affect fetal growth. Previous studies have examined these exposures individually, without considering the complexity of multiple concurrent exposures that women may face in the workplace, even though this reflects the true reality of occupational situations. The aim of this thesis was to study the effect of maternal occupational multi-exposures during pregnancy on fetal growth, using data from the Elfe mother-child cohort. The specific objectives were: (1) to identify profiles of maternal occupational multi-exposures to chemical, physical, biological, strenuous, organizational, and psychosocial factors during pregnancy, and to study their association with fetal growth; (2) to investigate the effect of maternal occupational multi-exposures on fetal growth using data-driven approaches; and (3) to examine the effect of selected occupational exposures, identified a priori, on fetal growth. We used data from the French national Elfe cohort. We characterized 47 occupational exposures using job-exposure matrices. Occupational exposure to each factor was defined using a multi-category variable, with different thresholds depending on the objectives. The health outcomes studied were small for gestational age (SGA), birthweight (BW), and head circumference (HC). Classification methods were used to determine multi-exposure profiles, for the first objective. Statistical data-driven methods such as Ewas, Lasso, and random forests were used to select variables addressing the second objective. Finally, an approach based on data from epidemiological and experimental literature was used to select the variables to be studied for the third objective. Among the 12,851 mothers included, women were exposed to a median of six factors. We identified four multi-exposure profiles: “low exposure, stress at work”, “strenuous, high organization, low decision”, “postural constraints, psychosocial factors”, “postural and strength constraints, chemical, and biological factors”. The profile “postural constraints, psychosocial factors” was associated with an increased risk of SGA and a decrease in HC among women who stopped working during the third trimester of pregnancy. Among the occupational exposures selected by data-driven methods, using a computer screen was associated with a decrease in BW, as was leaning forward or sideways in women not exposed to airborne germs. Task repetition was linked to a decrease in HC, as were oxygenated solvents – whether or not the women were exposed to airborne germs. Lastly, literature reviews enabled the selection of occupational exposures for which there were prior hypotheses regarding effects on fetal growth. In this approach, exposure to ultrafine particles was associated with a decrease in BW, and to oxygenated solvents with a decrease in HC. This multi-exposure approach, original in light of the international literature available to date, employed several types of statistical models to capture the complexity of the multi-exposures. Although replication is necessary, these studies contribute to a better understanding of the occupational exposome of pregnant women and may help improve occupational health prevention strategies.