Nouvelles approches de fusion reconstruction de données multi-modalités, basées sur des réseaux de neurones bayésiens : application à l'imagerie du sein et au Contrôle Non Destructif

This doctoral thesis is part of advanced research in non-destructive testing (NDT), developing innovative methodologies that leverage advances in artificial intelligence (AI) and machine learning to optimize inspection, detection, and characterization processes for structural defects in materials. The central objective of this study is to enhance the precision and reliability of imaging and diagnostic techniques by proposing hybrid approaches that integrate methods from signal processing, neural networks, and the physical modeling of the studied phenomena. First, a convolution-based approach was developed to model and approximate heat propagation in materials subjected to thermographic inspection. This methodology is particularly relevant for evaluating composite polymers, especially those affected by thin delamination defects, whose detection remains challenging due to low thermal contrasts and local conductivity variations. The ability of these models to exploit the spatial and temporal relationships inherent in thermographic data enables more precise reconstructions of internal anomalies. Furthermore, to improve the robustness and reliability of the obtained results, Bayesian neural networks (BNNs) were implemented to quantify the uncertainties associated with predictions. The integration of these probabilistic models serves as a crucial mechanism to enhance confidence in the decisions made by machine learning algorithms, particularly in scenarios where the available data is limited or partially noisy. This approach helps mitigate the risk of overfitting and improves the generalization of models to previously unseen configurations. Alongside these developments, this research explores the multimodal fusion of electromagnetic (EM) and ultrasonic (US) imaging techniques for breast lesion screening and diagnosis. The joint integration of these two modalities is based on leveraging the dielectric properties of electromagnetic waves, which provide information on tissue permittivity contrasts, and ultrasonic waves, which enable detailed structural reconstruction due to their high spatial resolution. To fully exploit the complementarity of these heterogeneous data sources, specialized convolutional neural networks (CNNs) have been developed, notably the Structurally-Aware Complex Cascaded Neural Network (SACCCNN), designed to maximize the preservation of underlying structures and ensure accurate reconstruction of both material and biological defects. Finally, in pursuit of performance enhancement and a reduction in the need for annotated data, transfer learning techniques have been explored, particularly Bayesian Bridge Transfer Learning (BBTL) and Bayesian Bridge Fused Transfer Learning (BBFTL). These methods allow for the efficient adaptation of pre-existing models from one imaging modality to another, thereby facilitating the generalization of algorithms to diverse configurations and strengthening their adaptability to real-world inspection and diagnostic scenarios. These contributions are part of an effort to advance the field of non-destructive testing and AIassisted medical imaging by developing more compact, precise, and robust tools while ensuring better interpretability of results and rigorous management of uncertainties associated with automated decision-making.