Master thesis : Matching ex vivo MRI scan and histological slices from multiple sclerosis patients

Abstract

In the context of Multiple sclerosis (MS) research, in order to detect the underlying cor- tical pathologies, we can rely on ultra-high-field (7T) quantitative magnetic resonance imaging (qMRI) as an alternative to the histological modality. Establishing a precise cor- relation between those 2 modalities can pave the way for improved in vivo detection and monitoring of the disease. Establishing this correspondence is very challenging due to the dimensionality mismatch between 3D MRI volumes and 2D histological sections, further compounded by severe non-linear tissue deformations such as shrinkage, stretching, and tearing introduced during chemical fixation and sections slicing. This master’s thesis presents the development of an automated, computational co-registration pipeline designed to accurately align 7T ex vivo MRI scans with corresponding histological slices. To manage the complex deformations, the framework relies on intermediate block- face imaging. The pipeline introduces a robust, multi-modal pre-processing stage featuring adaptive 3D brain masking and 2D colorimetric segmentation to strictly isolate cerebral tissue from background noise. The core 2D-to-3D affine registration is driven by a hybrid function that balances intensity-based Mutual Information (MI) with structural gradient matching. Optimized via a multi-resolution Z-axis grid search and Powell’s method, this approach successfully retrieves the arbitrary cutting plane within the 3D volume. Finally, a 2D-to-2D non-linear elastic registration, using Thin Plate Splines (TPS) and multi- resolution B-splines, is implemented to correct non-linear morphological distortions. The pipeline was evaluated on two post-mortem MS brain datasets, demonstrating a strong ability to recover the correct 3D cutting plane and to accurately align the outer cortical contours. Furthermore, the study highlights methodological key insights regarding multimodal optimization. We can mention the standard signal homogenization techniques such as N4 bias field correction were found to degrade the gradient-driven MI metric by flattening essential structural contrast. Additionally, while B-spline elastic deformations effectively matched outer cortical silhouettes, the lack of internal regularization anchors led to artificial warping of the internal cellular architecture. Finally, this work serves as a promising proof of concept, providing a standardized and reproducible mathematical framework for multimodal neuroimaging alignment. While the pipeline demonstrates excellent matching capabilities on high-quality datasets, it also high- lights the need for enhanced algorithmic robustness to handle lower-quality or artifact- heavy scans. By bridging the gap between engineering algorithms and neuropathological realities, it establishes a solid baseline for researchers and paves the way for future im- provements in histological segmentation.