Advancing brain MRI assessment using generative models

Abstract

Magnetic Resonance Imaging (MRI) has become an indispensable modality in neurological diagnostics, offering non-invasive, high-resolution insights into brain structure and function. Despite its clinical ubiquity, automated analysis of brain MRI faces persistent challenges, including domain variability across imaging sites, limited annotated data for target tasks, and the subtlety of certain developmental and pathological anomalies. Addressing these limitations is crucial for advancing diagnostic reliability, generalization, and scalability in real-world clinical settings. This thesis explores the integration of generative models into brain MRI analysis as a principled and label-efficient alternative to traditional discriminative frameworks. By modeling the underlying distribution of brain MRI data, generative approaches provide robust tools for learning from unlabeled data, simulating anatomical variability, and enhancing interpretability. We focus on three key areas of application: unsupervised MRI harmonization, unsupervised brain anomaly detection, and neonatal brain age estimation. First, we introduce Harmonizing Flows, a novel framework based on normalizing flows for unsupervised and source-free harmonization of multi-site MRI scans. This method effectively aligns data distributions across scanners while preserving clinically relevant features, significantly improving model generalization, even on unseen domains. Second, we leverage a progression of generative models and propose three unsupervised anomaly detection approaches, MAD-AD, DeCo-Diff, and REFLECT, each building upon the strengths and addressing the shortcomings of the previous, leading to progressively more robust and effective solutions. In particular, these models learn the manifold of healthy brain anatomy and isolate pathological deviations without requiring annotated anomalies, demonstrating strong performance across various datasets. Lastly, we propose a learning-based framework for predicting neonatal brain age, enabling the identification of infants at risk of neurodevelopmental delays by quantifying maturational discrepancies not evident in conventional structural assessments. Together, these contributions establish a cohesive, generative-model-based framework for brain MRI assessment that is scalable, interpretable, and clinically meaningful. Through extensive evaluation across diverse populations and imaging conditions, the proposed methods demonstrate improved robustness, enhanced diagnostic capability, and cross-domain generalizability. This work underscores the transformative potential of generative models in neuroimaging and paves the way toward more accessible, equitable, and effective brain health diagnostics.

Date	10 Sept 2025
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	José Dolz (Supervisor), Christian Desrosiers (Co-supervisor) & Gregory A. Lodygensky (Co-supervisor)