Variational Multi-Task Models for Image Analysis: Applications to Magnetic Resonance Imaging

Abstract: This thesis deals with the study and development of several variational multi-task models for solving inverse problems in imaging, with a particular focus on Magnetic Resonance Imaging (MRI). In most image processing problems, one usually deals with the reconstruction task, i.e., the task of reconstructing an image from indirect measurements, and then performs various operations, one after the other (i.e. sequentially), to improve the quality of the reconstruction and to extract useful information. However, recent developments in a variational context, have shown that performing those tasks jointly (i.e. in a multi-task framework) offers great benefits, and this is the perspective that we follow in this thesis. We go beyond traditional sequential approaches and set a new basis for variational multi-task methods for MRI analysis. We demonstrate that by sharing representation between tasks and carefully interconnecting them, one can create synergies across challenging problems and reduce error propagation. More precisely, firstly we propose a multi-task variational model to tackle the problems of image reconstruction and image segmentation using non-convex Bregman iteration. We describe theoretical and numerical details of the problem and its optimisation scheme. Moreover, we show that our multi-task model achieves better results in several examples and MRI applications than existing approaches in the same context. Secondly, we show that our approach can be extended to a multi-task reconstruction and segmentation model for the nonlinear inverse problem of velocity-encoded MRI. In this context, the aim is to estimate not only the magnitude from MRI data, but also the phase and its flow information, whilst simultaneously identify regions of interest through the segmentation task. Finally, we go beyond two-task frameworks and introduce for the first time a variational multi-task model to handle three imaging tasks. To this end, we design a variational multi-task framework addressing reconstruction, super-resolution and registration for improving the quality of MRI reconstruction. We demonstrate that our model is theoretically well-motivated and it outperforms sequential models whilst requiring less computational cost. Furthermore, we show through experimental results the potential of this approach for clinical applications.