- author = "Kevin Cortacero",
- title = "Programmation Genetique Cartesienne pour la Segmentation d'Images Biomedicales",
- school = "Institut Universitaire du Cancer de Toulouse",
- year = "2023",
- address = "France",
- month = "8 " # nov,
- keywords = "genetic algorithms, genetic programming, Cartesian Genetic Programming, Kartezio, Image processing, Explainable Artificial Intelligence, XAI, Machine Learning, Instance Segmentation",
-
URL = "
https://theses.hal.science/tel-04747703",
-
URL = "https://www.theses.fr/2023TOU30313",
-
URL = "
https://theses.hal.science/tel-04747703v1/file/2023TOU30313a.pdf",
- size = "266 pages",
-
abstract = "Over the past decade, Artificial Intelligence (AI),
and in particular Deep Learning (DL), has experienced
exponential growth. Although the main focus has been on
improving the accuracy of models, explainability,
frugality and dataset size reduction are fundamental
challenges that need to be addressed with increasing
urgency. While precision remains paramount, this thesis
research proposes a rebalancing of these metrics, in
order to achieve strong model performance in the
context of full transparency and Few-Shot
learning.
This thesis focuses upon the field of biomedical imaging and the segmentation of biological entities using computer vision. Herein, we introduce Kartezio, a modular Cartesian Genetic Programming (CGP)-based computational strategy that produces transparent and interpretable image processing algorithms through the iterative assembly and parametrisation of functions. The pipelines thus created demonstrate comparable performance to state-of-the-art DL approaches on image segmentation tasks, while requiring drastically smaller training datasets. This Few-Shot Learning and frugal method confers tremendous flexibility, as demonstrated herein through a series of biomedical Use Cases. In addition, this thesis explores several approaches to extract meaningful information from the structure of the generated algorithms, offering strategies to direct optimization of the algorithm in the future. Finally, this study proposes extensions of the published work wherein unlimited types of data can be integrated into a single multimodal evolutionary approach (MAGE).
Through this detailed exploration, this work not only introduces a methodology but also offers an alternative or complement to dominant black-box DL paradigms. This suggests that, despite the complexity of state-of-the-art DL solutions, we may have previously underestimated frugal and transparent strategies such as CGP, particularly in the field of Image Processing. In biomedical imaging, where model predictions can have profound implications, the idea of using modern AI techniques should not be sidelined either due to computational or data constraints, or a lack of transparency. By showcasing CGP models through several practical applications, this thesis argues for the adoption of eXplainable Artificial Intelligence (XAI) in the biomedical domain.",
-
notes = "HAL Id : tel-04747703
Supervisors: Salvatore Valitutti and Sylvain Cussat-Blanc",
