Date depot: 12 avril 2023 Titre: Extensions of graph-neural networks for medical image analysis Directrice de thèse: Florence ROSSANT (LISITE) Encadrante : Patricia CONDE-CESPEDES (LISITE) Domaine scientifique: Sciences et technologies de l'information et de la communication Thématique CNRS : Intelligence artificielle Resumé: It is well-known that machine learning approaches, such as CNNs (convolutional neural networks) are often used in computer vision for medical image processing for tasks such as segmentation, object detection or image classification. A major limitation of the existing methods is that they rely on grid-like data; however, the structure of physiological recordings is often irregular and unordered which makes it difficult to treat them as matrix data. Furthermore, traditional CNNs (convolutional neural networks) do not capture complex neighborhood interaction as they analyze local areas based on fixed connectivity (determined by the convolutional operation), leading to limited performance and interpretability of the analysis of functional and anatomical structures. Besides, an image might contain data that suggests interactions between objects. In such situations, graph-based representation is the most suitable tool to tackle the input data. In simple words, graphs can be defined as a set of objects called nodes or vertices connected by links also called edges. For instance, we can mention the brain activity analysis used to represent either physical or functional connectivity across different brain regions. Graph neural networks (GNNs) are neural networks that operate over graphs. However, they take as input directly graph data manually defined. One challenge is the automatic definition of graph data starting directly from the image data, where the vertices correspond to the entities or ROIs specific to the problem (e.g. brain regions), and edges represent the connectivity of these entities. Another aspect to into account is the dynamic aspect of the graphs. Many real-world medical applications are dynamic in nature. In a graph context, this means their nodes, edges, and features can change over time. The purpose of this project is to extend GNNs to image data with applications in medicine.