Project
Organs like the lungs, liver, kidneys, and bowel, among others, are subject to respiratory motion, which has a large impact, during some types of medical treatment such as radiotherapy. The challenge of compensating for respiratory motion, which can lead to unintended radiation exposure of healthy tissues, remains a significant concern. To address this challenge, this project proposes geometric deep learning as a novel approach to improve motion modeling. Unlike traditional convolutional neural networks, geometric learning operates on graph structures, allowing them to capture complex relationships and deformations more effectively. The project aims to propose a model for modeling and reconstruction of organs and other anatomical structures. Ultimately, the project aims to enhance the effectiveness and objectivity of cancer diagnosis and treatment through cutting-edge technology and interdisciplinary collaboration.
Within the framework of this project, as a first step, we will present new segmentation methodologies based on graph neural networks and also on conventional approaches. As a main contribution, we will propose a deformable attention graph neural network - liver reconstruction method, which is an end-to-end trainable model that infers a 3D liver mesh structure at any time during treatment. Moreover, as a further step, we will introduce a spatio-temporal graph neural network for motion modeling. Finally, we will experiment with dose delivery approaches also based on graph neural networks. Based on the following we formulate three objectives of the project proposal: