Sequential and Active Learning for Advancing MRI Acquisition and Anomaly Detection in Brain Images

Abstract:
Magnetic Resonance Imaging (MRI) has transformed medical imaging, particularly in neuroscience, by enabling non-invasive, high-resolution 3D mapping of tissue properties for accurate diagnosis and surgical planning of brain pathologies. However, the increasing demand for faster MRI acquisition and the growing volume of high-dimensional brain MRI datasets to analyze pose significant challenges for model learning. Sequential learning provides a robust solution, serving as a cornerstone of stochastic gradient descent and enabling deep neural networks (DNNs) to efficiently process large-scale datasets. Although traditional machine learning algorithms struggle to scale with large volumes of high-dimensional data, they remain effective for specific tasks. For this reason, in the first part of this thesis, we investigate how traditional ML algorithms can be scaled by adopting sequential learning techniques. We focus specifically on adapting the Expectation-Maximization (EM) algorithm and mixture models to manage both large volumes and high-dimensional data. Our approach involves applying these scaled-up models to two practical challenges: (i) detecting subtle abnormalities in brain MRIs of de novo Parkinsonian patients, and (ii) solving the inverse problem of brain MRI "fingerprinting" reconstruction. We demonstrate that these models not only perform competitively with DNNs but are also computationally frugal. In the second part, we explore another perspective on sequential learning and computational efficiency, centered on the idea of learning while acquiring the minimal amount of data necessary to achieve a given task. We apply this concept to efficient MRI k-space sampling, leveraging diffusion models and Bayesian experimental design. Finally, both parts required the development of new software tools tailored for rapid iteration and experimentation with mixture models, the EM algorithm, and diffusion-based methods. These tools are introduced in the final part as two Python- and JAX-based packages—onlineEM and diffuse—designed to streamline research in these areas.

The thesis was supervised by:

- Florence Forbes, Research Director, Inria

- Michel Dojat, Research Director, Inserm, Institute of neuroscience of Grenoble

- Carole Lartizien, Research Director, CNRS, CREATIS

Members of the jury:

- Jeffrey Fessler, Full Professor, University of Michigan, Reviewer
- Julia Schnabel, Full Professor, Technical University of Munich, Reviewer
- Isabelle Bloch, Full Professor, Sorbonne Université, Examiner
- Carole Sudre, Full Professor, University College London, Examiner
- Olivier Detante, Full Professor, MD, Université Grenoble Alpes, CHU Grenoble Alpes, President