Résumé
High serous ovarian cancer is the most lethal gynecologic malignancy, characterized by late-stage diagnosis and a five-year survival rate near 50%. Improving outcomes requires diagnostic strategies that leverage complementary information from imaging, histopathology, and molecular analyses. High-field Magnetic Resonance Imaging (MRI) is crucial in this effort due to its capacity to produce high-resolution images, with the objective of linking the high-fidelity MRI signal to the nature of the underlying tumor tissues. Multimodal learning is critical in this context, enabling the integration of heterogeneous data. However, for these models to be biologically valid, features from Whole Slide Imaging (WSI)—such as cellular morphology—must align precisely with their anatomical location in Magnetic Resonance Imaging (MRI). Without strict spatial coherence, biological signals are misassigned, compromising the robustness and interpretability of downstream deep learning models.Consequently, image registration is essential to achieve the voxel-level alignment required for multimodal analysis. This is inherently challenging: MRI is a smooth, low-resolution 3D modality, whereas Histo is high-resolution 2D. Additionally, Histo data entails nonlinear geometric distortions from tissue processing (fixation, slicing). These differences necessitate robust pipelines combining rigid, affine, and complex non-rigid transformations.This work introduces a systematic and robust comparative framework to benchmark multimodal image registration methods, focusing on the alignment of high-resolution Histo whole-slide images with volumetric MRI. Accurate spatial integration serves as the foundation for transformative downstream applications, including the development of deep learning models for virtual biopsies that predict Histo-specific features directly from non-invasive MRI, as well as the design of multimodal models that integrate MRI with spatially aligned Histo data to improve diagnostic accuracy and prognostic assessment.We leverage a dataset of approximately matched pairs of MRI 9.4T and Whole Slide Images, H&E-stained images from surgically resected tumors. To ensure spatial fidelity, specimens were scanned ex vivo using a 3D printed mold that preserves tissue orientation. A single 2D tissue section was prepared for histological analysis, and the corresponding slice was extracted from the volumetric MRI. Each input pair—2D MRI and Histo —provides cross-scale, modality-diverse views with minimal out-of-plane mismatch.A benchmarking framework was designed to evaluate four registration paradigms for non-rigid tissue deformation correction: feature-based methods relying on anatomical landmark matching; intensity-based methods optimizing similarity metrics like mutual information or cross-correlation; deformation-based models such as B-splines and thin-plate splines for smooth, local warping; and deep learning methods including VoxelMorph and MRI-histology networks. All input pairs are pre-processed with intensity normalization and spatial cropping to mitigate modality-specific photometric variability.Because ex vivo processing introduces anatomical and physical variability, perfect alignment is unattainable. We therefore aim to identify the pipeline that yields the most robust and biologically meaningful alignment. Performance is assessed using Mutual Information (MI), Normalized Mutual Information (NMI), Normalized Cross-Correlation (NCC), and Structural Similarity Index (SSIM). While the evaluation is ongoing, current results show that the Rigid + Affine + B-spline pipeline achieves the best overall performance (MAE = 0.1652, 1-SSIM = 0.8982, 1-(NCC+1)/2 = 0.4441). These findings suggest that a multi-stage registration approach best preserves spatial fidelity.