par Galvez Jiménez, Laura 
Président du jury Bontempi, Gianluca
Promoteur Decaestecker, Christine
Co-Promoteur Debeir, Olivier
Publication Non publié, 2025-12-15
Président du jury Bontempi, Gianluca
Promoteur Decaestecker, Christine
Co-Promoteur Debeir, Olivier
Publication Non publié, 2025-12-15
Thèse de doctorat
| Résumé : | Digital pathology has rapidly advanced with the integration of AI, which is increasingly used to support pathologists in complex image analyses, enhancing diagnostic efficiency, reducing workload, and enabling the detection of subtle features that even experts may elude.Despite their potential, the development of computational pathology faces critical challenges that limit the trust in AI. Histopathological images are inherently complex, large and costly to annotate, leading to limited availability of datasets with potential noisy, inconsistent, and incomplete labels. Addressing these limitations requires the development of deep learning models capable of handling imperfect annotations while maintaining the high performance standards demanded in medical applications.This thesis investigates the development of robust deep learning techniques for digital pathology in the presence of imperfect annotations. The primary focus is on tasks that integrate detection, segmentation, and classification of cells and other tissue structures across multiple organs. We begin with an overview of digital pathology and the role of deep learning in this domain, then examine key challenges and propose corresponding solutions.We address imperfect annotations through different strategies and setups. First, we assess the impact of annotation errors in instance segmentation and classification of cells in histopathology images and use data cleaning techniques to improve the training set. We then analyze the effects of imperfect annotations on the combined tasks of detection, segmentation, and classification and investigate the conditions for determining an appropriate number of training epochs to prevent overfitting to annotation noise during training. Finally, we explore advanced approaches that integrate semi-supervised and active learning to address missing or incomplete annotations, with a particular focus on the segmentation of larger structures. Together, these methods enable robust models that perform well despite noisy or incomplete labels, reducing reliance on exhaustive manual annotation.Annotator variability further challenges model reliability, as pathologists may produce inconsistent labels. We propose a probabilistic framework to assess annotator behavior based on the concept of self-consistency and to estimate consensus ground truth that accounts for annotator reliability and variability. Furthermore, we compare this approach with alternative probabilistic soft-labeling methods.Overall, this research highlights the important role of high-quality annotations in digital pathology while demonstrating that carefully designed computational strategies can effectively mitigate challenges introduced by noisy, incomplete, or inconsistent labels. By addressing both the technical and practical aspects of annotation imperfections, this thesis contributes to enhancing the reliability and robustness of deep learning models in digital pathology. |
