Improving the accuracy of cortical surface registration while preserving individual characteristics is a difficult task, since cortical shape varies significantly across different individuals. In this paper we propose an extension of Multimodal Surface Matching (MSM) towards a hierarchical groupwise framework to address these challenges. Central to the success of this framework has been re-implementation of MSM to improve its efficiency. Experiments show that the proposed method supports unbiased and efficient alignment of large groups of up to 250 subjects, while significantly improving the alignment of cortical folds, relative to alignment to a single population average template. This implies significant advantages with regards to the search for subtle cortical signatures of cognition and disease. The code can be found at https://github. com/rbesenczi/newMSM.
2D to 3D registration is essential in tasks such as diagnosis, surgical navigation, environmental understanding, navigation in robotics, autonomous systems, or augmented reality. In medical imaging, the aim is often to place a 2D image in a 3D volumetric observation to augment or compare information. Current approaches for rigid single slice in volume registration are limited by requirements such as pose initialization, stacks of adjacent slices, or reliable anatomical landmarks. Here, we propose a self-supervised 2D/3D registration approach to match a single 2D slice to the corresponding 3D volume. The method does not require anatomical priors or an initialization. It addresses the dimensionality disparity and establishes correspondences between 2D in-plane and 3D out-of-plane rotation-equivariant features by using group equivariant CNNs. These rotation-equivariant features are extracted from the 2D query slice and aligned with their 3D counterparts. Results demonstrate the robustness of the proposed slice-in-volume registration on the NSCLC-Radiomics CT and KIRBY21 MRI datasets, attaining an absolute median angle error of less than 2 ^∘ C and a mean-matching feature accuracy of 89
Deep learning models have revolutionized image registration but their accuracy can degrade under unforeseen data variations (domain shifts). It is crucial to assess model robustness under such shifts, often accomplished using simulated domain shifts and expert annotations, e.g., landmarks. This work presents ProactiV-Reg, an annotation-free approach that utilizes a learnable image mapping: it iteratively adjusts a moving image to align with a fixed image under simulated domain shifts. The distances between the perturbed and the optimized images reveal model robustness. We evaluate ProactiV-Reg on three models, showcasing its ability to detect robustness differences, identify dominant perturbations, and provide insights into the model's input requirements.
Investigating the maturation process of the brain through clinical studies is crucial for advancing our understanding of the developing brain. Neuroimaging studies employing magnetic resonance images (MRI) often necessitate pre-alignment of images to a standardized space. Registration of diffusion tensor images (DTI) has the potential to better align white matter (WM) structures than using structural MRI only, as it enables alignment of fiber orientation at each voxel. However, microstructural data are not well suited to accurately register the cortical gray matter (cGM) ribbon. In spite of this, in many studies the alignment of subjects is primarily guided by a single modality. In this work, we propose a multi-channel attention-based deep learning registration approach that selects the most salient features from multiple image modalities to improve alignment of individual MR images to a common atlas space. We apply the technique to align multi-channel datasets composed of structural T2-weighted (T2w) MRI and DTI maps into atlas space. The quantitative and qualitative evaluation confirmed that when we use the two modalities we obtain good alignment of anatomical structures, while also improving the alignment of the underlying white matter tracts. Moreover, we show that learning a spatially-varying attention map for weighting the two modalities obtains superior performance than the baseline multi-channel method that does not incorporate attention.
We present a new model for deformable image registration, which learns in an unsupervised way a data-specific similarity metric. The proposed method consists of two neural networks, one that maps pairs of input images to transformations which align them, and one that provides the similarity metric whose maximisation guides the image alignment. We parametrise the similarity metric as an energy-based model, which is simple to train and allows us to improve the accuracy of image registration compared to other models with learnt similarity metrics by taking advantage of a more general mathematical formulation, as well as larger datasets. We also achieve substantial improvement in the accuracy of inter-patient image registration on MRI scans from the OASIS dataset compared to models that rely on traditional functions.
