The FastModelAlign module in SlicerMorph is for 3D model registration using a unified workflow approach. It performs scaling, rigid (RANSAC and ICP), affine, and deformable registration steps. Users can select any combination of registration steps via checkboxes to create custom registration pipelines. The rigid registration involves RANSAC and ICP steps from downsampled pointclouds derived from 3D models. Affine and deformable registrations use Coherent Point Drift (CPD) algorithms. For more technical implementation details, see the related ALPACA automated landmarking publication.

We provided two partial models of the same mountain beaver skull derived from two separate photogrammetric runs as sample data: partial_photogram_model1.obj is a partial skull with no bottom, and partial_photogram_model2.obj is a partial skull with no top. After installing the SlicerMorph extension, they can be downloaded from the Sample data module: clicking the partial photogrammetry models button under the FastModelAlign tab will download the two sample models to the Slicer cache folder and load them into Slicer automatically (also downloadable here: SlicerMorph sample data. The following tutorial are based on these two sample models.

If using the Sample Data module, the two partial models should be loaded into Slicer automatically. If downloaded manually, drag and drop the downloaded sample models into 3D Slicer. The two models have distinct sizes because no size reference was provided for the two photogrammetric runs.

Switch to the FastModelAlign module. Select partial_photogram_model2 from the Source Model drop-down menu and partial_photogram_model1 from the Target Model drop-down menu. Also create a model from the Output registered model drop-down menu. In this tutorial, the output registered model is named as test_registration.

The module now uses a unified workflow with four registration step options:

• Scaling: Scale the source model to approximate the size of the target model
• Rigid: Apply rigid registration (rotation and translation) using RANSAC and ICP
• Affine: Apply affine transformation using CPD (allows shearing and non-uniform scaling)
• **Deformaegistration After the source and target models are selected and at least one registration step is checked, the Testing pointcloud subsampling and Run Registration buttons are enabled.

(Optional) Click Testing pointcloud subsampling button to preview the density of the downsampled source and target point clouds. Adjust the Point density adjustment slider to control the number of points in each point cloud. Usually, 4,000 to 6,000 points per point cloud would be optimal.

Click the Run Registration button to execute the selected registration steps in sequence. For example, if both Scaling and Rigid are checked, the process first scales the source model to approximate the size of the target model, then rigidly registers the scaled source to the target. After the process is finished, the user-specified output model test_registration (red) is generated and represents the source model transformed according to the selected registration steps.

(Optional) Click Testing pointcloud subsampling button can show the density of the downsampled source and target point clouds. Adjusting the Point density adjustment bar to adjust the number of points in each point cloud. Usually, 4,000 to 6,000 points per point cloud would be optimal.

Click the Run rigid registration button to rigidly register the source partial_photogram_model2 to the target model partial_photogram_model1. If Skip scaling is unchecked, the process first scale the source model to approximate the size of the target model, then rigidly registered the source to the target.After the process is finished, the user specified output model test_registration (red) is generated and represents the source model rigidly registered to the target model (yellow).

After running registration, the module generates transform nodes based on the selected steps:

• Scaling transform: Named as [source_model_name]_scaling (if scaling was selected)
• Rigid transform: Named as [source_model_name]_rigid (if rigid was selected)
• Affine transform: Named as [source_model_name]_affine (if affine was selected)
• Deformable transform: Named as [source_model_name]_deformable (if deformable was selected, and fast option is unchecked)

The transform nodes are organized hierarchically according to the registration sequence. For example:

• If Scaling and Rigid are selected: The scaling node is nested under the rigid transformation node
• If Scaling, Rigid, and Affine are selected: The scaling node is under the rigid node, which is under the affine node
• If Deformable is included: All previous transforms are nested under the deformable transform node

This hierarchical structure ensures transforms are applied in the correct sequence: Scaling → Rigid → Affine → Deformable.

To repeat and visualize the transformation process, users can drag the original source model partial_photogram_model2.obj (blue) to the appropriate transform node in the Transform hierarchy.

Switch to the Transform hierarchy tab to visualize the hierarchical transformation of the source model node through the selected registration steps.egistration node.

Switch to the Transform hierarchy, users can visualize the hierarchical transformation of the source model node, which is first scaled by the scaling node, then rigidly registered by the rigid transformation node. Affine registration When the Affine checkbox is selected, an affine transformation is applied using Coherent Point Drift (CPD) algorithm. The affine transformation allows for shearing and non-uniform scaling in addition to rotation and translation. This step is useful when the source and target models have different proportions or slight distortions.

The affine transformation is applied to the previously registered model (after any selected scaling and/or rigid steps) and directly modifies the model vertices. An affine transform node is created for reference and to maintain the transform hierarchy

After running a rigid transformation, the Run affine registration button will be enabled. Click this button will perform an affine deformable transformation for the scaled (if Scaling optin is left unchecked) and rigidly registered source model. Therefore, the user defined output model (test_registration (red)) is then undergone an affine transformation to match with the target (yellow).

When Fast Mode is enabled, the CPD deformation field is applied directly to the model vertices using Radial Basis Function (RBF) interpolation with thin plate spline kernel. This mode is faster and produces a deformed model node named [source_model_name]_deformed (displayed in green). The deformation is hardened into the model vertices, and no transform node is created.

When Fast Mode is disabled, a grid transform is created that defines a smooth deformation field. This mode:

• Creates a vtkMRMLGridTransformNode named [source_model_name]_deformable
• Produces a hardened deformed model named [source_model_name]_deformed
• The grid resolution is controlled by the Grid spacing slider in Advanced Settings (1 = finest ~256³ grid, 10 = coarsest ~64³ grid)
• Uses RBF interpolation to create a smooth displacement field on a regular grid

The deformed model (green) represents the final registered result with local shape variations matching the target (yellow).

Key parameters for deformable registration (adjustable in Advanced Settings):

• Alpha: Controls the smoothness of the deformation (higher = smoother, default: 2.0)
• Beta: Controls the regularization of the deformation (higher = less deformation, default: 2.0)
• CPD Iterations: Maximum iterations for CPD optimization (default: 100)
• CPD Tolerance: Convergence tolerance for CPD (default: 0.001)
• Grid Spacing: Controls grid density for grid transform mode (1-10, only for grid mode)

To repeat and visualize the full registration sequence, users can drag the original untransformed source model partial_photogram_model2.obj (blue) into the appropriate transform node for the complete sequence of transformations

To repeat and visualize the affine registration, users can drag the original untransformed source model partial_photogram_model2.obj (blue) into the scaling node partial_photogram_model2_scaling for a full sequence of scaling, rigid registration, and affine transformation.

Users can adjust the parameter settings under the Advanced Settings tab before running registration. The parameters are organized by registration type:

• Point Density: Controls the density of subsampled point clouds (default: 1.0)
• Poisson Subsample: If checked, uses Poisson disk sampling for more uniform point distribution
• Normal Search Radius: Radius for normal estimation (default: 2.0)

For detailed information about rigid registration parameters, please refer to the ALPACA publication.

• FPFH Search Radius: Search radius for Fast Point Feature Histogram (FPFH) features (default: 5.0)
• FPFH Neighbors: Number of neighbors for FPFH computation (default: 100)
• Maximum CPD Threshold: Distance threshold for correspondence rejection (default: 3.0)
• Max RANSAC Iterations: Maximum RANSAC iterations (default: 1,000,000)
• ICP Distance Threshold: Distance threshold for Iterative Closest Point refinement (default: 1.5)

• Alpha: Smoothness parameter for CPD deformation (default: 2.0)
  • Higher values produce smoother deformations
  • Lower values allow more local variation
• Beta: Regularization parameter for CPD (default: 2.0)
  • Higher values reduce the amount of deformation
  • Lower values allow more aggressive deformation
• CPD Iterations: Maximum iterations for CPD optimization (default: 100)
• CPD Tolerance: Convergence tolerance for stopping criterion (default: 0.001)
• Grid Spacing: Controls grid resolution for grid transform mode (1-10)
  • 1 = finest resolution (~256 grid points per dimension)
  • 10 = coarsest resolution (~64 grid points per dimension)
  • Only applicable when Fast Mode is disabled
• Fast Mode: When enabled, applies deformation directly to vertices (faster, no grid transform created)