Spatial Registration

The most fundamental registration approach uses fiducial markers or tissue landmarks to compute an affine transformation matrix. This matrix applies translation, rotation, scaling, and shearing to map a moving image onto a fixed reference. - Control Points: Manually or automatically identified corresponding points (e.g., tissue folds, H&E-stained features) - Transformation Model: Solves for the optimal rigid or affine parameters minimizing the distance between point pairs - Application: Essential for aligning consecutive tissue sections where global tissue morphology is preserved but orientation differs

Biological tissues are elastic and undergo non-linear deformations during sectioning and mounting. Deformable registration uses B-splines, thin-plate splines, or diffeomorphic models to warp local regions independently. - Free-Form Deformation: A grid of control points is optimized to maximize local similarity - Diffeomorphic Registration: Ensures the transformation is smooth, invertible, and topology-preserving, critical for tracking anatomical structures - Use Case: Correcting for tissue tears, folds, or differential shrinkage between serial sections

Automated registration algorithms iteratively optimize a similarity metric that quantifies alignment quality between a moving and fixed image. The choice of metric depends on the imaging modality. - Normalized Cross-Correlation (NCC): Robust to linear intensity scaling, ideal for aligning H&E-stained sections - Mutual Information (MI): A statistical measure that captures non-linear intensity relationships, essential for cross-modality registration (e.g., aligning fluorescence microscopy to H&E) - Mean Squared Difference (MSD): Suitable only when images share identical intensity distributions

The ultimate goal of spatial registration is to integrate data from disparate sources into a common coordinate system. This enables the direct comparison of spatial transcriptomics with immunofluorescence or H&E histology. - Sequential Section Alignment: Registering a transcriptomic capture spot array to a high-resolution microscopy image of an adjacent section - Cross-Technology Mapping: Aligning 10x Visium spots to MERFISH single-cell data for resolution enhancement - Challenge: Overcoming fundamentally different feature spaces where pixel intensities do not correlate directly

Instead of raw pixel intensities, this method extracts scale-invariant features (e.g., SIFT, SURF) or geometric shapes (cell boundaries, nuclei centroids) to drive registration. - Keypoint Matching: Identifies distinctive local image patches that are robust to rotation and scale changes - Geometric Consensus: Uses RANSAC (Random Sample Consensus) to filter out incorrect matches and estimate a robust transformation - Advantage: Highly efficient for large whole-slide images and robust to partial tissue occlusion

Aligning spatial transcriptomics data presents unique challenges: the capture spot array must be mapped onto a tissue image. This requires specialized pipelines. - Fiducial Frame Detection: Automatically detecting the fluorescent fiducial border printed around the capture area to define the initial coordinate system - Spot-to-Pixel Mapping: Transforming the grid of gene expression vectors to overlay precisely on the histological image - Tool Example: STUtility and STAlign are R packages designed specifically for registering and integrating multiple 10x Visium samples into a common spatial framework

A statistical measure of the degree to which a variable's values at nearby locations are more similar than expected by chance. After registration, calculating Moran's I on gene expression validates that biological spatial patterns are preserved rather than disrupted by the alignment transformation.

• Moran's I ranges from -1 (dispersed) to +1 (clustered)
• Used to identify spatially variable genes (SVGs)
• Validates registration accuracy on known marker genes

The unsupervised identification of tissue regions with coherent gene expression profiles and histology. Accurate registration is a prerequisite for domain detection across serial sections, allowing algorithms like spatial hidden Markov models or graph-based clustering to track anatomical structures consistently.

• Uses spatial neighborhood graphs
• Identifies functional tissue compartments
• Depends on precise inter-section alignment

The computational process of partitioning a digital tissue image into distinct anatomical or functional regions based on pixel-level classification. Deep learning models like U-Net perform segmentation on registered images to define boundaries that guide downstream spatial analysis.

• Distinguishes cortex, medulla, and tumor margins
• Provides masks for region-specific differential expression
• Often the reference layer for multi-modal registration

A computational method for removing technical variation between multiple spatial transcriptomic samples while preserving true biological heterogeneity. After registration aligns tissue sections, batch correction algorithms like Harmony or MNN adjust for slide-to-slide variability without erasing genuine spatial signals.

• Essential for multi-sample atlases
• Prevents technical artifacts from being misinterpreted as biology
• Applied post-registration in integrated data objects

A computational method that orders cells based on their spatial coordinates and gene expression profiles to reconstruct dynamic processes like differentiation in situ. Registration across time-series or depth-series sections enables pseudotime algorithms to map developmental lineages in physical space.

• Reconstructs developmental gradients
• Links gene expression changes to spatial position
• Requires temporally ordered, registered sections