image-matchmaker documentation
Image-Matchmaker is a framework that leverages segmentation masks to achieve alignment between two volumetric datasets of different modalities.
Image-Matchmaker is designed to optimize a transformation that maps a moving image/volume to align with a fixed image/volume.
Moving Volume: The source dataset that is spatially transformed (warped) to match the target.
Fixed Volume: The stationary reference dataset that defines the target coordinate space for the registration.
Overview of the Image-Matchmaker registration workflow.
The workflow begins by taking two segmentation masks from different modalities (e.g., EM and LM) as the primary inputs. Then these masks enter a sequential alignment pipeline to progressively refine the spatial correspondence between the volumes.
The pipeline consists of the following steps:
Pre-alignment (SVD): The masks undergo an initial global alignment using Singular Value Decomposition (SVD). This provides a coarse starting point by aligning the centroids and principal axes of the segmented structures.
Rigid Registration (Elastix): Following pre-alignment, a rigid transformation is performed using the Elastix toolbox to account for basic rotation and translation differences.
Coherent Point Drift (CPD) Registration: To further refine the alignment, CPD registration is applied to the point clouds derived from the rigid registration results using probreg, allowing for more nuanced local adjustments.
Feature Matching: The workflow then identifies specific correspondences (e.g., matching individual nuclei) between the two volumes to establish a set of definitive landmarks shared by both datasets. Three matching algorithms are available: a sparse integer linear program (
ilp, via cvxpy), an optimal one-to-one assignment (hungarian), and an entropy-regularized soft assignment (sinkhorn).B-Spline Registration (Elastix): The final deformable alignment is performed using a B-spline transformation.
Note: Crucially, the B-Spline Registration step utilizes the original segmentation masks and the established matching landmarks as direct inputs, rather than relying on the intermediate results from the CPD stage
Final Transformation Application
The result of the B-spline registration is a sequential transformation comprising three refined stages: rigid, rough B-spline, and fine B-spline. This sequence can be applied directly to the target channels of the original volumes (e.g., raw EM or fluorescence LM) to bring them into a unified coordinate space.
Get started: see Installation to set up the environment, then follow the Quick Start to run your first registration.